1fio(1) General Commands Manual fio(1)
2
6 fio - flexible I/O tester
7
9 fio [options] [jobfile]...
10
12 fio is a tool that will spawn a number of threads or processes doing a
13 particular type of I/O action as specified by the user. The typical
14 use of fio is to write a job file matching the I/O load one wants to
15 simulate.
16
18 --debug=type
19 Enable verbose tracing type of various fio actions. May be `all'
20 for all types or individual types separated by a comma (e.g.
21 `--debug=file,mem' will enable file and memory debugging).
22 `help' will list all available tracing options.
23 --parse-only
25 Parse options only, don't start any I/O.
26 --merge-blktrace-only
28 Merge blktraces only, don't start any I/O.
29 --output=filename
31 Write output to filename.
32 --output-format=format
34 Set the reporting format to `normal', `terse', `json', or
35 `json+'. Multiple formats can be selected, separate by a comma.
36 `terse' is a CSV based format. `json+' is like `json', except it
37 adds a full dump of the latency buckets.
38 --bandwidth-log
40 Generate aggregate bandwidth logs.
41 --minimal
43 Print statistics in a terse, semicolon-delimited format.
44 --append-terse
46 Print statistics in selected mode AND terse, semicolon-delimited
47 format. Deprecated, use --output-format instead to select mul‐
48 tiple formats.
49 --terse-version=version
51 Set terse version output format (default `3', or `2', `4', `5').
52 --version
54 Print version information and exit.
55 --help Print a summary of the command line options and exit.
57 --cpuclock-test
59 Perform test and validation of internal CPU clock.
60 --crctest=[test]
62 Test the speed of the built-in checksumming functions. If no ar‐
63 gument is given, all of them are tested. Alternatively, a comma
64 separated list can be passed, in which case the given ones are
65 tested.
66 --cmdhelp=command
68 Print help information for command. May be `all' for all com‐
69 mands.
70 --enghelp=[ioengine[,command]]
72 List all commands defined by ioengine, or print help for command
73 defined by ioengine. If no ioengine is given, list all available
74 ioengines.
75 --showcmd=jobfile
77 Convert jobfile to a set of command-line options.
78 --readonly
80 Turn on safety read-only checks, preventing writes and trims.
81 The --readonly option is an extra safety guard to prevent users
82 from accidentally starting a write or trim workload when that is
83 not desired. Fio will only modify the device under test if
84 `rw=write/randwrite/rw/randrw/trim/randtrim/trimwrite' is given.
85 This safety net can be used as an extra precaution.
86 --eta=when
88 Specifies when real-time ETA estimate should be printed. when
89 may be `always', `never' or `auto'. `auto' is the default, it
90 prints ETA when requested if the output is a TTY. `always' dis‐
91 regards the output type, and prints ETA when requested. `never'
92 never prints ETA.
93 --eta-interval=time
95 By default, fio requests client ETA status roughly every second.
96 With this option, the interval is configurable. Fio imposes a
97 minimum allowed time to avoid flooding the console, less than
98 250 msec is not supported.
99 --eta-newline=time
101 Force a new line for every time period passed. When the unit is
102 omitted, the value is interpreted in seconds.
103 --status-interval=time
105 Force a full status dump of cumulative (from job start) values
106 at time intervals. This option does *not* provide per-period
107 measurements. So values such as bandwidth are running averages.
108 When the time unit is omitted, time is interpreted in seconds.
109 Note that using this option with `--output-format=json' will
110 yield output that technically isn't valid json, since the output
111 will be collated sets of valid json. It will need to be split
112 into valid sets of json after the run.
113 --section=name
115 Only run specified section name in job file. Multiple sections
116 can be specified. The --section option allows one to combine
117 related jobs into one file. E.g. one job file could define
118 light, moderate, and heavy sections. Tell fio to run only the
119 "heavy" section by giving `--section=heavy' command line option.
120 One can also specify the "write" operations in one section and
121 "verify" operation in another section. The --section option only
122 applies to job sections. The reserved *global* section is always
123 parsed and used.
124 --alloc-size=kb
126 Allocate additional internal smalloc pools of size kb in KiB.
127 The --alloc-size option increases shared memory set aside for
128 use by fio. If running large jobs with randommap enabled, fio
129 can run out of memory. Smalloc is an internal allocator for
130 shared structures from a fixed size memory pool and can grow to
131 16 pools. The pool size defaults to 16MiB. NOTE: While running
132 `.fio_smalloc.*' backing store files are visible in `/tmp'.
133 --warnings-fatal
135 All fio parser warnings are fatal, causing fio to exit with an
136 error.
137 --max-jobs=nr
139 Set the maximum number of threads/processes to support to nr.
140 NOTE: On Linux, it may be necessary to increase the shared-mem‐
141 ory limit (`/proc/sys/kernel/shmmax') if fio runs into errors
142 while creating jobs.
143 --server=args
145 Start a backend server, with args specifying what to listen to.
146 See CLIENT/SERVER section.
147 --daemonize=pidfile
149 Background a fio server, writing the pid to the given pidfile
150 file.
151 --client=hostname
153 Instead of running the jobs locally, send and run them on the
154 given hostname or set of hostnames. See CLIENT/SERVER section.
155 --remote-config=file
157 Tell fio server to load this local file.
158 --idle-prof=option
160 Report CPU idleness. option is one of the following:
161 calibrate
163 Run unit work calibration only and exit.
164 system Show aggregate system idleness and unit work.
166 percpu As system but also show per CPU idleness.
168 --inflate-log=log
170 Inflate and output compressed log.
171 --trigger-file=file
173 Execute trigger command when file exists.
174 --trigger-timeout=time
176 Execute trigger at this time.
177 --trigger=command
179 Set this command as local trigger.
180 --trigger-remote=command
182 Set this command as remote trigger.
183 --aux-path=path
185 Use the directory specified by path for generated state files
186 instead of the current working directory.
187
189 Any parameters following the options will be assumed to be job files,
190 unless they match a job file parameter. Multiple job files can be
191 listed and each job file will be regarded as a separate group. Fio will
192 stonewall execution between each group.
193 Fio accepts one or more job files describing what it is supposed to do.
195 The job file format is the classic ini file, where the names enclosed
196 in [] brackets define the job name. You are free to use any ASCII name
197 you want, except *global* which has special meaning. Following the job
198 name is a sequence of zero or more parameters, one per line, that de‐
199 fine the behavior of the job. If the first character in a line is a ';'
200 or a '#', the entire line is discarded as a comment.
201 A *global* section sets defaults for the jobs described in that file. A
203 job may override a *global* section parameter, and a job file may even
204 have several *global* sections if so desired. A job is only affected by
205 a *global* section residing above it.
206 The --cmdhelp option also lists all options. If used with an command
208 argument, --cmdhelp will detail the given command.
209 See the `examples/' directory for inspiration on how to write job
211 files. Note the copyright and license requirements currently apply to
212 `examples/' files.
213 Note that the maximum length of a line in the job file is 8192 bytes.
215
217 Some parameters take an option of a given type, such as an integer or a
218 string. Anywhere a numeric value is required, an arithmetic expression
219 may be used, provided it is surrounded by parentheses. Supported opera‐
220 tors are:
221 addition (+)
223 subtraction (-)
225 multiplication (*)
227 division (/)
229 modulus (%)
231 exponentiation (^)
233 For time values in expressions, units are microseconds by default. This
235 is different than for time values not in expressions (not enclosed in
236 parentheses).
237
239 The following parameter types are used.
240 str String. A sequence of alphanumeric characters.
242 time Integer with possible time suffix. Without a unit value is in‐
244 terpreted as seconds unless otherwise specified. Accepts a suf‐
245 fix of 'd' for days, 'h' for hours, 'm' for minutes, 's' for
246 seconds, 'ms' (or 'msec') for milliseconds and 'us' (or 'usec')
247 for microseconds. For example, use 10m for 10 minutes.
248 int Integer. A whole number value, which may contain an integer pre‐
250 fix and an integer suffix.
251 [*integer prefix*] **number** [*integer suffix*]
253 The optional *integer prefix* specifies the number's base. The
255 default is decimal. *0x* specifies hexadecimal.
256 The optional *integer suffix* specifies the number's units, and
258 includes an optional unit prefix and an optional unit. For quan‐
259 tities of data, the default unit is bytes. For quantities of
260 time, the default unit is seconds unless otherwise specified.
261 With `kb_base=1000', fio follows international standards for
263 unit prefixes. To specify power-of-10 decimal values defined in
264 the International System of Units (SI):
265 K means kilo (K) or 1000
267 M means mega (M) or 1000**2
268 G means giga (G) or 1000**3
269 T means tera (T) or 1000**4
270 P means peta (P) or 1000**5
271 To specify power-of-2 binary values defined in IEC 80000-13:
273 Ki means kibi (Ki) or 1024
275 Mi means mebi (Mi) or 1024**2
276 Gi means gibi (Gi) or 1024**3
277 Ti means tebi (Ti) or 1024**4
278 Pi means pebi (Pi) or 1024**5
279 For Zone Block Device Mode:
281 z means Zone
283 With `kb_base=1024' (the default), the unit prefixes are oppo‐
284 site from those specified in the SI and IEC 80000-13 standards
285 to provide compatibility with old scripts. For example, 4k means
286 4096.
287 For quantities of data, an optional unit of 'B' may be included
289 (e.g., 'kB' is the same as 'k').
290 The *integer suffix* is not case sensitive (e.g., m/mi mean
292 mebi/mega, not milli). 'b' and 'B' both mean byte, not bit.
293 Examples with `kb_base=1000':
295 4 KiB: 4096, 4096b, 4096B, 4k, 4kb, 4kB, 4K, 4KB
297 1 MiB: 1048576, 1m, 1024k
298 1 MB: 1000000, 1mi, 1000ki
299 1 TiB: 1073741824, 1t, 1024m, 1048576k
300 1 TB: 1000000000, 1ti, 1000mi, 1000000ki
301 Examples with `kb_base=1024' (default):
303 4 KiB: 4096, 4096b, 4096B, 4k, 4kb, 4kB, 4K, 4KB
305 1 MiB: 1048576, 1m, 1024k
306 1 MB: 1000000, 1mi, 1000ki
307 1 TiB: 1073741824, 1t, 1024m, 1048576k
308 1 TB: 1000000000, 1ti, 1000mi, 1000000ki
309 To specify times (units are not case sensitive):
311 D means days
313 H means hours
314 M mean minutes
315 s or sec means seconds (default)
316 ms or msec means milliseconds
317 us or usec means microseconds
318 `z' suffix specifies that the value is measured in zones. Value
320 is recalculated once block device's zone size becomes known.
321 If the option accepts an upper and lower range, use a colon ':'
323 or minus '-' to separate such values. See irange parameter type.
324 If the lower value specified happens to be larger than the upper
325 value the two values are swapped.
326 bool Boolean. Usually parsed as an integer, however only defined for
328 true and false (1 and 0).
329 irange Integer range with suffix. Allows value range to be given, such
331 as 1024-4096. A colon may also be used as the separator, e.g.
332 1k:4k. If the option allows two sets of ranges, they can be
333 specified with a ',' or '/' delimiter: 1k-4k/8k-32k. Also see
334 int parameter type.
335 float_list
337 A list of floating point numbers, separated by a ':' character.
338
340 With the above in mind, here follows the complete list of fio job pa‐
341 rameters.
342 Units
344 kb_base=int
345 Select the interpretation of unit prefixes in input parameters.
346 1000 Inputs comply with IEC 80000-13 and the Interna‐
348 tional System of Units (SI). Use:
349 - power-of-2 values with IEC prefixes (e.g., KiB)
351 - power-of-10 values with SI prefixes (e.g., kB)
352 1024 Compatibility mode (default). To avoid breaking
354 old scripts:
355 - power-of-2 values with SI prefixes
357 - power-of-10 values with IEC prefixes
358 See bs for more details on input parameters.
360 Outputs always use correct prefixes. Most outputs include both
362 side-by-side, like:
363 bw=2383.3kB/s (2327.4KiB/s)
365 If only one value is reported, then kb_base selects the one to
367 use:
368 1000 -- SI prefixes
370 1024 -- IEC prefixes
371 unit_base=int
373 Base unit for reporting. Allowed values are:
374 0 Use auto-detection (default).
376 8 Byte based.
378 1 Bit based.
380 Job description
382 name=str
383 ASCII name of the job. This may be used to override the name
384 printed by fio for this job. Otherwise the job name is used. On
385 the command line this parameter has the special purpose of also
386 signaling the start of a new job.
387 description=str
389 Text description of the job. Doesn't do anything except dump
390 this text description when this job is run. It's not parsed.
391 loops=int
393 Run the specified number of iterations of this job. Used to re‐
394 peat the same workload a given number of times. Defaults to 1.
395 numjobs=int
397 Create the specified number of clones of this job. Each clone of
398 job is spawned as an independent thread or process. May be used
399 to setup a larger number of threads/processes doing the same
400 thing. Each thread is reported separately; to see statistics for
401 all clones as a whole, use group_reporting in conjunction with
402 new_group. See --max-jobs. Default: 1.
403 Time related parameters
405 runtime=time
406 Tell fio to terminate processing after the specified period of
407 time. It can be quite hard to determine for how long a specified
408 job will run, so this parameter is handy to cap the total run‐
409 time to a given time. When the unit is omitted, the value is in‐
410 terpreted in seconds.
411 time_based
413 If set, fio will run for the duration of the runtime specified
414 even if the file(s) are completely read or written. It will sim‐
415 ply loop over the same workload as many times as the runtime al‐
416 lows.
417 startdelay=irange(int)
419 Delay the start of job for the specified amount of time. Can be
420 a single value or a range. When given as a range, each thread
421 will choose a value randomly from within the range. Value is in
422 seconds if a unit is omitted.
423 ramp_time=time
425 If set, fio will run the specified workload for this amount of
426 time before logging any performance numbers. Useful for letting
427 performance settle before logging results, thus minimizing the
428 runtime required for stable results. Note that the ramp_time is
429 considered lead in time for a job, thus it will increase the to‐
430 tal runtime if a special timeout or runtime is specified. When
431 the unit is omitted, the value is given in seconds.
432 clocksource=str
434 Use the given clocksource as the base of timing. The supported
435 options are:
436 gettimeofday
438 gettimeofday(2)
439 clock_gettime
441 clock_gettime(2)
442 cpu Internal CPU clock source
444 cpu is the preferred clocksource if it is reliable, as it is
446 very fast (and fio is heavy on time calls). Fio will automati‐
447 cally use this clocksource if it's supported and considered re‐
448 liable on the system it is running on, unless another clock‐
449 source is specifically set. For x86/x86-64 CPUs, this means sup‐
450 porting TSC Invariant.
451 gtod_reduce=bool
453 Enable all of the gettimeofday(2) reducing options (dis‐
454 able_clat, disable_slat, disable_bw_measurement) plus reduce
455 precision of the timeout somewhat to really shrink the gettime‐
456 ofday(2) call count. With this option enabled, we only do about
457 0.4% of the gettimeofday(2) calls we would have done if all time
458 keeping was enabled.
459 gtod_cpu=int
461 Sometimes it's cheaper to dedicate a single thread of execution
462 to just getting the current time. Fio (and databases, for in‐
463 stance) are very intensive on gettimeofday(2) calls. With this
464 option, you can set one CPU aside for doing nothing but logging
465 current time to a shared memory location. Then the other
466 threads/processes that run I/O workloads need only copy that
467 segment, instead of entering the kernel with a gettimeofday(2)
468 call. The CPU set aside for doing these time calls will be ex‐
469 cluded from other uses. Fio will manually clear it from the CPU
470 mask of other jobs.
471 Target file/device
473 directory=str
474 Prefix filenames with this directory. Used to place files in a
475 different location than `./'. You can specify a number of direc‐
476 tories by separating the names with a ':' character. These di‐
477 rectories will be assigned equally distributed to job clones
478 created by numjobs as long as they are using generated file‐
479 names. If specific filename(s) are set fio will use the first
480 listed directory, and thereby matching the filename semantic
481 (which generates a file for each clone if not specified, but
482 lets all clones use the same file if set).
483 See the filename option for information on how to escape ':'
485 characters within the directory path itself.
486 Note: To control the directory fio will use for internal state
488 files use --aux-path.
489 filename=str
491 Fio normally makes up a filename based on the job name, thread
492 number, and file number (see filename_format). If you want to
493 share files between threads in a job or several jobs with fixed
494 file paths, specify a filename for each of them to override the
495 default. If the ioengine is file based, you can specify a number
496 of files by separating the names with a ':' colon. So if you
497 wanted a job to open `/dev/sda' and `/dev/sdb' as the two work‐
498 ing files, you would use `filename=/dev/sda:/dev/sdb'. This also
499 means that whenever this option is specified, nrfiles is ig‐
500 nored. The size of regular files specified by this option will
501 be size divided by number of files unless an explicit size is
502 specified by filesize.
503 Each colon in the wanted path must be escaped with a '\' charac‐
505 ter. For instance, if the path is `/dev/dsk/foo@3,0:c' then you
506 would use `filename=/dev/dsk/foo@3,0\:c' and if the path is
507 `F:\filename' then you would use `filename=F\:\filename'.
508 On Windows, disk devices are accessed as `\\.\PhysicalDrive0'
510 for the first device, `\\.\PhysicalDrive1' for the second etc.
511 Note: Windows and FreeBSD prevent write access to areas of the
512 disk containing in-use data (e.g. filesystems).
513 The filename `-' is a reserved name, meaning *stdin* or *std‐
515 out*. Which of the two depends on the read/write direction set.
516 filename_format=str
518 If sharing multiple files between jobs, it is usually necessary
519 to have fio generate the exact names that you want. By default,
520 fio will name a file based on the default file format specifica‐
521 tion of `jobname.jobnumber.filenumber'. With this option, that
522 can be customized. Fio will recognize and replace the following
523 keywords in this string:
524 $jobname
526 The name of the worker thread or process.
527 $clientuid
529 IP of the fio process when using client/server
530 mode.
531 $jobnum
533 The incremental number of the worker thread or
534 process.
535 $filenum
537 The incremental number of the file for that worker
538 thread or process.
539 To have dependent jobs share a set of files, this option can be
541 set to have fio generate filenames that are shared between the
542 two. For instance, if `testfiles.$filenum' is specified, file
543 number 4 for any job will be named `testfiles.4'. The default of
544 `$jobname.$jobnum.$filenum' will be used if no other format
545 specifier is given.
546 If you specify a path then the directories will be created up to
548 the main directory for the file. So for example if you specify
549 `a/b/c/$jobnum` then the directories a/b/c will be created be‐
550 fore the file setup part of the job. If you specify directory
551 then the path will be relative that directory, otherwise it is
552 treated as the absolute path.
553 unique_filename=bool
555 To avoid collisions between networked clients, fio defaults to
556 prefixing any generated filenames (with a directory specified)
557 with the source of the client connecting. To disable this behav‐
558 ior, set this option to 0.
559 opendir=str
561 Recursively open any files below directory str.
562 lockfile=str
564 Fio defaults to not locking any files before it does I/O to
565 them. If a file or file descriptor is shared, fio can serialize
566 I/O to that file to make the end result consistent. This is
567 usual for emulating real workloads that share files. The lock
568 modes are:
569 none No locking. The default.
571 exclusive
573 Only one thread or process may do I/O at a time,
574 excluding all others.
575 readwrite
577 Read-write locking on the file. Many readers may
578 access the file at the same time, but writes get
579 exclusive access.
580 nrfiles=int
582 Number of files to use for this job. Defaults to 1. The size of
583 files will be size divided by this unless explicit size is spec‐
584 ified by filesize. Files are created for each thread separately,
585 and each file will have a file number within its name by de‐
586 fault, as explained in filename section.
587 openfiles=int
589 Number of files to keep open at the same time. Defaults to the
590 same as nrfiles, can be set smaller to limit the number simulta‐
591 neous opens.
592 file_service_type=str
594 Defines how fio decides which file from a job to service next.
595 The following types are defined:
596 random Choose a file at random.
598 roundrobin
600 Round robin over opened files. This is the de‐
601 fault.
602 sequential
604 Finish one file before moving on to the next. Mul‐
605 tiple files can still be open depending on open‐
606 files.
607 zipf Use a Zipf distribution to decide what file to ac‐
609 cess.
610 pareto Use a Pareto distribution to decide what file to
612 access.
613 normal Use a Gaussian (normal) distribution to decide
615 what file to access.
616 gauss Alias for normal.
618 For random, roundrobin, and sequential, a postfix can be ap‐
620 pended to tell fio how many I/Os to issue before switching to a
621 new file. For example, specifying `file_service_type=random:8'
622 would cause fio to issue 8 I/Os before selecting a new file at
623 random. For the non-uniform distributions, a floating point
624 postfix can be given to influence how the distribution is
625 skewed. See random_distribution for a description of how that
626 would work.
627 ioscheduler=str
629 Attempt to switch the device hosting the file to the specified
630 I/O scheduler before running. If the file is a pipe, a character
631 device file or if device hosting the file could not be deter‐
632 mined, this option is ignored.
633 create_serialize=bool
635 If true, serialize the file creation for the jobs. This may be
636 handy to avoid interleaving of data files, which may greatly de‐
637 pend on the filesystem used and even the number of processors in
638 the system. Default: true.
639 create_fsync=bool
641 fsync(2) the data file after creation. This is the default.
642 create_on_open=bool
644 If true, don't pre-create files but allow the job's open() to
645 create a file when it's time to do I/O. Default: false -- pre-
646 create all necessary files when the job starts.
647 create_only=bool
649 If true, fio will only run the setup phase of the job. If files
650 need to be laid out or updated on disk, only that will be done
651 -- the actual job contents are not executed. Default: false.
652 allow_file_create=bool
654 If true, fio is permitted to create files as part of its work‐
655 load. If this option is false, then fio will error out if the
656 files it needs to use don't already exist. Default: true.
657 allow_mounted_write=bool
659 If this isn't set, fio will abort jobs that are destructive
660 (e.g. that write) to what appears to be a mounted device or par‐
661 tition. This should help catch creating inadvertently destruc‐
662 tive tests, not realizing that the test will destroy data on the
663 mounted file system. Note that some platforms don't allow writ‐
664 ing against a mounted device regardless of this option. Default:
665 false.
666 pre_read=bool
668 If this is given, files will be pre-read into memory before
669 starting the given I/O operation. This will also clear the in‐
670 validate flag, since it is pointless to pre-read and then drop
671 the cache. This will only work for I/O engines that are seek-
672 able, since they allow you to read the same data multiple times.
673 Thus it will not work on non-seekable I/O engines (e.g. network,
674 splice). Default: false.
675 unlink=bool
677 Unlink the job files when done. Not the default, as repeated
678 runs of that job would then waste time recreating the file set
679 again and again. Default: false.
680 unlink_each_loop=bool
682 Unlink job files after each iteration or loop. Default: false.
683 zonemode=str
685 Accepted values are:
686 none The zonerange, zonesize zonecapacity and zoneskip
688 parameters are ignored.
689 strided
691 I/O happens in a single zone until zonesize bytes
692 have been transferred. After that number of bytes
693 has been transferred processing of the next zone
694 starts. The zonecapacity parameter is ignored.
695 zbd Zoned block device mode. I/O happens sequentially
697 in each zone, even if random I/O has been se‐
698 lected. Random I/O happens across all zones in‐
699 stead of being restricted to a single zone. Trim
700 is handled using a zone reset operation. Trim only
701 considers non-empty sequential write required and
702 sequential write preferred zones.
703 zonerange=int
705 For zonemode=strided, this is the size of a single zone. See
706 also zonesize and zoneskip.
707 For zonemode=zbd, this parameter is ignored.
709 zonesize=int
711 For zonemode=strided, this is the number of bytes to transfer
712 before skipping zoneskip bytes. If this parameter is smaller
713 than zonerange then only a fraction of each zone with zonerange
714 bytes will be accessed. If this parameter is larger than zon‐
715 erange then each zone will be accessed multiple times before
716 skipping to the next zone.
717 For zonemode=zbd, this is the size of a single zone. The zon‐
719 erange parameter is ignored in this mode. For a job accessing a
720 zoned block device, the specified zonesize must be 0 or equal to
721 the device zone size. For a regular block device or file, the
722 specified zonesize must be at least 512B.
723 zonecapacity=int
725 For zonemode=zbd, this defines the capacity of a single zone,
726 which is the accessible area starting from the zone start ad‐
727 dress. This parameter only applies when using zonemode=zbd in
728 combination with regular block devices. If not specified it de‐
729 faults to the zone size. If the target device is a zoned block
730 device, the zone capacity is obtained from the device informa‐
731 tion and this option is ignored.
732 zoneskip=int[z]
734 For zonemode=strided, the number of bytes to skip after zonesize
735 bytes of data have been transferred.
736 For zonemode=zbd, the zonesize aligned number of bytes to skip
738 once a zone is fully written (write workloads) or all written
739 data in the zone have been read (read workloads). This parameter
740 is valid only for sequential workloads and ignored for random
741 workloads. For read workloads, see also read_beyond_wp.
742 read_beyond_wp=bool
745 This parameter applies to zonemode=zbd only.
746 Zoned block devices are block devices that consist of multiple
748 zones. Each zone has a type, e.g. conventional or sequential. A
749 conventional zone can be written at any offset that is a multi‐
750 ple of the block size. Sequential zones must be written sequen‐
751 tially. The position at which a write must occur is called the
752 write pointer. A zoned block device can be either host managed
753 or host aware. For host managed devices the host must ensure
754 that writes happen sequentially. Fio recognizes host managed de‐
755 vices and serializes writes to sequential zones for these de‐
756 vices.
757 If a read occurs in a sequential zone beyond the write pointer
759 then the zoned block device will complete the read without read‐
760 ing any data from the storage medium. Since such reads lead to
761 unrealistically high bandwidth and IOPS numbers fio only reads
762 beyond the write pointer if explicitly told to do so. Default:
763 false.
764 max_open_zones=int
766 When running a random write test across an entire drive many
767 more zones will be open than in a typical application workload.
768 Hence this command line option that allows to limit the number
769 of open zones. The number of open zones is defined as the number
770 of zones to which write commands are issued by all threads/pro‐
771 cesses.
772 job_max_open_zones=int
774 Limit on the number of simultaneously opened zones per single
775 thread/process.
776 ignore_zone_limits=bool
778 If this option is used, fio will ignore the maximum number of
779 open zones limit of the zoned block device in use, thus allowing
780 the option max_open_zones value to be larger than the device re‐
781 ported limit. Default: false.
782 zone_reset_threshold=float
784 A number between zero and one that indicates the ratio of logi‐
785 cal blocks with data to the total number of logical blocks in
786 the test above which zones should be reset periodically.
787 zone_reset_frequency=float
789 A number between zero and one that indicates how often a zone
790 reset should be issued if the zone reset threshold has been ex‐
791 ceeded. A zone reset is submitted after each (1 / zone_re‐
792 set_frequency) write requests. This and the previous parameter
793 can be used to simulate garbage collection activity.
794 I/O type
797 direct=bool
798 If value is true, use non-buffered I/O. This is usually O_DI‐
799 RECT. Note that OpenBSD and ZFS on Solaris don't support direct
800 I/O. On Windows the synchronous ioengines don't support direct
801 I/O. Default: false.
802 atomic=bool
804 If value is true, attempt to use atomic direct I/O. Atomic
805 writes are guaranteed to be stable once acknowledged by the op‐
806 erating system. Only Linux supports O_ATOMIC right now.
807 buffered=bool
809 If value is true, use buffered I/O. This is the opposite of the
810 direct option. Defaults to true.
811 readwrite=str, rw=str
813 Type of I/O pattern. Accepted values are:
814 read Sequential reads.
816 write Sequential writes.
818 trim Sequential trims (Linux block devices and SCSI
820 character devices only).
821 randread
823 Random reads.
824 randwrite
826 Random writes.
827 randtrim
829 Random trims (Linux block devices and SCSI charac‐
830 ter devices only).
831 rw,readwrite
833 Sequential mixed reads and writes.
834 randrw Random mixed reads and writes.
836 trimwrite
838 Sequential trim+write sequences. Blocks will be
839 trimmed first, then the same blocks will be writ‐
840 ten to.
841 Fio defaults to read if the option is not specified. For the
843 mixed I/O types, the default is to split them 50/50. For certain
844 types of I/O the result may still be skewed a bit, since the
845 speed may be different.
846 It is possible to specify the number of I/Os to do before get‐
848 ting a new offset by appending `:<nr>' to the end of the string
849 given. For a random read, it would look like `rw=randread:8' for
850 passing in an offset modifier with a value of 8. If the suffix
851 is used with a sequential I/O pattern, then the `<nr>' value
852 specified will be added to the generated offset for each I/O
853 turning sequential I/O into sequential I/O with holes. For in‐
854 stance, using `rw=write:4k' will skip 4k for every write. Also
855 see the rw_sequencer option.
856 rw_sequencer=str
858 If an offset modifier is given by appending a number to the
859 `rw=str' line, then this option controls how that number modi‐
860 fies the I/O offset being generated. Accepted values are:
861 sequential
863 Generate sequential offset.
864 identical
866 Generate the same offset.
867 sequential is only useful for random I/O, where fio would nor‐
869 mally generate a new random offset for every I/O. If you append
870 e.g. 8 to randread, you would get a new random offset for every
871 8 I/Os. The result would be a seek for only every 8 I/Os, in‐
872 stead of for every I/O. Use `rw=randread:8' to specify that. As
873 sequential I/O is already sequential, setting sequential for
874 that would not result in any differences. identical behaves in a
875 similar fashion, except it sends the same offset 8 number of
876 times before generating a new offset.
877 unified_rw_reporting=str
879 Fio normally reports statistics on a per data direction basis,
880 meaning that reads, writes, and trims are accounted and reported
881 separately. This option determines whether fio reports the re‐
882 sults normally, summed together, or as both options. Accepted
883 values are:
884 none Normal statistics reporting.
886 mixed Statistics are summed per data direction and reported to‐
888 gether.
889 both Statistics are reported normally, followed by the mixed
891 statistics.
892 0 Backward-compatible alias for none.
894 1 Backward-compatible alias for mixed.
896 2 Alias for both.
898 randrepeat=bool
900 Seed the random number generator used for random I/O patterns in
901 a predictable way so the pattern is repeatable across runs. De‐
902 fault: true.
903 allrandrepeat=bool
905 Seed all random number generators in a predictable way so re‐
906 sults are repeatable across runs. Default: false.
907 randseed=int
909 Seed the random number generators based on this seed value, to
910 be able to control what sequence of output is being generated.
911 If not set, the random sequence depends on the randrepeat set‐
912 ting.
913 fallocate=str
915 Whether pre-allocation is performed when laying down files. Ac‐
916 cepted values are:
917 none Do not pre-allocate space.
919 native Use a platform's native pre-allocation call but
921 fall back to none behavior if it fails/is not im‐
922 plemented.
923 posix Pre-allocate via posix_fallocate(3).
925 keep Pre-allocate via fallocate(2) with FAL‐
927 LOC_FL_KEEP_SIZE set.
928 truncate
930 Extend file to final size using ftruncate|(2) in‐
931 stead of allocating.
932 0 Backward-compatible alias for none.
934 1 Backward-compatible alias for posix.
936 May not be available on all supported platforms. keep is only
938 available on Linux. If using ZFS on Solaris this cannot be set
939 to posix because ZFS doesn't support pre-allocation. Default:
940 native if any pre-allocation methods except truncate are avail‐
941 able, none if not.
942 Note that using truncate on Windows will interact surprisingly
944 with non-sequential write patterns. When writing to a file that
945 has been extended by setting the end-of-file information, Win‐
946 dows will backfill the unwritten portion of the file up to that
947 offset with zeroes before issuing the new write. This means that
948 a single small write to the end of an extended file will stall
949 until the entire file has been filled with zeroes.
950 fadvise_hint=str
952 Use posix_fadvise(2) or posix_madvise(2) to advise the kernel
953 what I/O patterns are likely to be issued. Accepted values are:
954 0 Backwards compatible hint for "no hint".
956 1 Backwards compatible hint for "advise with fio
958 workload type". This uses FADV_RANDOM for a random
959 workload, and FADV_SEQUENTIAL for a sequential
960 workload.
961 sequential
963 Advise using FADV_SEQUENTIAL.
964 random Advise using FADV_RANDOM.
966 write_hint=str
968 Use fcntl(2) to advise the kernel what life time to expect from
969 a write. Only supported on Linux, as of version 4.13. Accepted
970 values are:
971 none No particular life time associated with this file.
973 short Data written to this file has a short life time.
975 medium Data written to this file has a medium life time.
977 long Data written to this file has a long life time.
979 extreme
981 Data written to this file has a very long life
982 time.
983 The values are all relative to each other, and no absolute mean‐
985 ing should be associated with them.
986 offset=int[%|z]
988 Start I/O at the provided offset in the file, given as either a
989 fixed size in bytes, zones or a percentage. If a percentage is
990 given, the generated offset will be aligned to the minimum
991 blocksize or to the value of offset_align if provided. Data be‐
992 fore the given offset will not be touched. This effectively caps
993 the file size at `real_size - offset'. Can be combined with size
994 to constrain the start and end range of the I/O workload. A
995 percentage can be specified by a number between 1 and 100 fol‐
996 lowed by '%', for example, `offset=20%' to specify 20%. In ZBD
997 mode, value can be set as number of zones using 'z'.
998 offset_align=int
1000 If set to non-zero value, the byte offset generated by a per‐
1001 centage offset is aligned upwards to this value. Defaults to 0
1002 meaning that a percentage offset is aligned to the minimum block
1003 size.
1004 offset_increment=int[%|z]
1006 If this is provided, then the real offset becomes `offset + off‐
1007 set_increment * thread_number', where the thread number is a
1008 counter that starts at 0 and is incremented for each sub-job
1009 (i.e. when numjobs option is specified). This option is useful
1010 if there are several jobs which are intended to operate on a
1011 file in parallel disjoint segments, with even spacing between
1012 the starting points. Percentages can be used for this option.
1013 If a percentage is given, the generated offset will be aligned
1014 to the minimum blocksize or to the value of offset_align if pro‐
1015 vided.In ZBD mode, value can be set as number of zones using
1016 'z'.
1017 number_ios=int
1019 Fio will normally perform I/Os until it has exhausted the size
1020 of the region set by size, or if it exhaust the allocated time
1021 (or hits an error condition). With this setting, the range/size
1022 can be set independently of the number of I/Os to perform. When
1023 fio reaches this number, it will exit normally and report sta‐
1024 tus. Note that this does not extend the amount of I/O that will
1025 be done, it will only stop fio if this condition is met before
1026 other end-of-job criteria.
1027 fsync=int
1029 If writing to a file, issue an fsync(2) (or its equivalent) of
1030 the dirty data for every number of blocks given. For example, if
1031 you give 32 as a parameter, fio will sync the file after every
1032 32 writes issued. If fio is using non-buffered I/O, we may not
1033 sync the file. The exception is the sg I/O engine, which syn‐
1034 chronizes the disk cache anyway. Defaults to 0, which means fio
1035 does not periodically issue and wait for a sync to complete.
1036 Also see end_fsync and fsync_on_close.
1037 fdatasync=int
1039 Like fsync but uses fdatasync(2) to only sync data and not meta‐
1040 data blocks. In Windows, FreeBSD, DragonFlyBSD or OSX there is
1041 no fdatasync(2) so this falls back to using fsync(2). Defaults
1042 to 0, which means fio does not periodically issue and wait for a
1043 data-only sync to complete.
1044 write_barrier=int
1046 Make every N-th write a barrier write.
1047 sync_file_range=str:int
1049 Use sync_file_range(2) for every int number of write operations.
1050 Fio will track range of writes that have happened since the last
1051 sync_file_range(2) call. str can currently be one or more of:
1052 wait_before
1054 SYNC_FILE_RANGE_WAIT_BEFORE
1055 write SYNC_FILE_RANGE_WRITE
1057 wait_after
1059 SYNC_FILE_RANGE_WRITE_AFTER
1060 So if you do `sync_file_range=wait_before,write:8', fio would
1062 use `SYNC_FILE_RANGE_WAIT_BEFORE | SYNC_FILE_RANGE_WRITE' for
1063 every 8 writes. Also see the sync_file_range(2) man page. This
1064 option is Linux specific.
1065 overwrite=bool
1067 If true, writes to a file will always overwrite existing data.
1068 If the file doesn't already exist, it will be created before the
1069 write phase begins. If the file exists and is large enough for
1070 the specified write phase, nothing will be done. Default: false.
1071 end_fsync=bool
1073 If true, fsync(2) file contents when a write stage has com‐
1074 pleted. Default: false.
1075 fsync_on_close=bool
1077 If true, fio will fsync(2) a dirty file on close. This differs
1078 from end_fsync in that it will happen on every file close, not
1079 just at the end of the job. Default: false.
1080 rwmixread=int
1082 Percentage of a mixed workload that should be reads. Default:
1083 50.
1084 rwmixwrite=int
1086 Percentage of a mixed workload that should be writes. If both
1087 rwmixread and rwmixwrite is given and the values do not add up
1088 to 100%, the latter of the two will be used to override the
1089 first. This may interfere with a given rate setting, if fio is
1090 asked to limit reads or writes to a certain rate. If that is the
1091 case, then the distribution may be skewed. Default: 50.
1092 random_distribution=str:float[:float][,str:float][,str:float]
1094 By default, fio will use a completely uniform random distribu‐
1095 tion when asked to perform random I/O. Sometimes it is useful to
1096 skew the distribution in specific ways, ensuring that some parts
1097 of the data is more hot than others. fio includes the following
1098 distribution models:
1099 random Uniform random distribution
1101 zipf Zipf distribution
1103 pareto Pareto distribution
1105 normal Normal (Gaussian) distribution
1107 zoned Zoned random distribution zoned_abs Zoned absolute
1109 random distribution
1110 When using a zipf or pareto distribution, an input value is also
1112 needed to define the access pattern. For zipf, this is the `Zipf
1113 theta'. For pareto, it's the `Pareto power'. Fio includes a
1114 test program, fio-genzipf, that can be used visualize what the
1115 given input values will yield in terms of hit rates. If you
1116 wanted to use zipf with a `theta' of 1.2, you would use `ran‐
1117 dom_distribution=zipf:1.2' as the option. If a non-uniform model
1118 is used, fio will disable use of the random map. For the normal
1119 distribution, a normal (Gaussian) deviation is supplied as a
1120 value between 0 and 100.
1121 The second, optional float is allowed for pareto, zipf and nor‐
1123 mal distributions. It allows to set base of distribution in non-
1124 default place, giving more control over most probable outcome.
1125 This value is in range [0-1] which maps linearly to range of
1126 possible random values. Defaults are: random for pareto and
1127 zipf, and 0.5 for normal. If you wanted to use zipf with a
1128 `theta` of 1.2 centered on 1/4 of allowed value range, you would
1129 use `random_distibution=zipf:1.2:0.25`.
1130 For a zoned distribution, fio supports specifying percentages of
1132 I/O access that should fall within what range of the file or de‐
1133 vice. For example, given a criteria of:
1134 60% of accesses should be to the first 10%
1136 30% of accesses should be to the next 20%
1137 8% of accesses should be to the next 30%
1138 2% of accesses should be to the next 40%
1139 we can define that through zoning of the random accesses. For
1141 the above example, the user would do:
1142 random_distribution=zoned:60/10:30/20:8/30:2/40
1144 A zoned_abs distribution works exactly like thezoned, except
1146 that it takes absolute sizes. For example, let's say you wanted
1147 to define access according to the following criteria:
1148 60% of accesses should be to the first 20G
1150 30% of accesses should be to the next 100G
1151 10% of accesses should be to the next 500G
1152 we can define an absolute zoning distribution with:
1154 random_distribution=zoned:60/10:30/20:8/30:2/40
1156 For both zoned and zoned_abs, fio supports defining up to 256
1158 separate zones.
1159 Similarly to how bssplit works for setting ranges and percent‐
1161 ages of block sizes. Like bssplit, it's possible to specify sep‐
1162 arate zones for reads, writes, and trims. If just one set is
1163 given, it'll apply to all of them.
1164 percentage_random=int[,int][,int]
1166 For a random workload, set how big a percentage should be ran‐
1167 dom. This defaults to 100%, in which case the workload is fully
1168 random. It can be set from anywhere from 0 to 100. Setting it to
1169 0 would make the workload fully sequential. Any setting in be‐
1170 tween will result in a random mix of sequential and random I/O,
1171 at the given percentages. Comma-separated values may be speci‐
1172 fied for reads, writes, and trims as described in blocksize.
1173 norandommap
1175 Normally fio will cover every block of the file when doing ran‐
1176 dom I/O. If this option is given, fio will just get a new random
1177 offset without looking at past I/O history. This means that some
1178 blocks may not be read or written, and that some blocks may be
1179 read/written more than once. If this option is used with verify
1180 and multiple blocksizes (via bsrange), only intact blocks are
1181 verified, i.e., partially-overwritten blocks are ignored. With
1182 an async I/O engine and an I/O depth > 1, it is possible for the
1183 same block to be overwritten, which can cause verification er‐
1184 rors. Either do not use norandommap in this case, or also use
1185 the lfsr random generator.
1186 softrandommap=bool
1188 See norandommap. If fio runs with the random block map enabled
1189 and it fails to allocate the map, if this option is set it will
1190 continue without a random block map. As coverage will not be as
1191 complete as with random maps, this option is disabled by de‐
1192 fault.
1193 random_generator=str
1195 Fio supports the following engines for generating I/O offsets
1196 for random I/O:
1197 tausworthe
1199 Strong 2^88 cycle random number generator.
1200 lfsr Linear feedback shift register generator.
1202 tausworthe64
1204 Strong 64-bit 2^258 cycle random number generator.
1205 tausworthe is a strong random number generator, but it requires
1207 tracking on the side if we want to ensure that blocks are only
1208 read or written once. lfsr guarantees that we never generate the
1209 same offset twice, and it's also less computationally expensive.
1210 It's not a true random generator, however, though for I/O pur‐
1211 poses it's typically good enough. lfsr only works with single
1212 block sizes, not with workloads that use multiple block sizes.
1213 If used with such a workload, fio may read or write some blocks
1214 multiple times. The default value is tausworthe, unless the re‐
1215 quired space exceeds 2^32 blocks. If it does, then tausworthe64
1216 is selected automatically.
1217 Block size
1219 blocksize=int[,int][,int], bs=int[,int][,int]
1220 The block size in bytes used for I/O units. Default: 4096. A
1221 single value applies to reads, writes, and trims. Comma-sepa‐
1222 rated values may be specified for reads, writes, and trims. A
1223 value not terminated in a comma applies to subsequent types. Ex‐
1224 amples:
1225 bs=256k means 256k for reads, writes and trims.
1227 bs=8k,32k means 8k for reads, 32k for writes and
1228 trims.
1229 bs=8k,32k, means 8k for reads, 32k for writes, and
1230 default for trims.
1231 bs=,8k means default for reads, 8k for writes and
1232 trims.
1233 bs=,8k, means default for reads, 8k for writes,
1234 and default for trims.
1235 blocksize_range=irange[,irange][,irange],
1237 bsrange=irange[,irange][,irange]
1238 A range of block sizes in bytes for I/O units. The issued I/O
1239 unit will always be a multiple of the minimum size, unless
1240 blocksize_unaligned is set. Comma-separated ranges may be spec‐
1241 ified for reads, writes, and trims as described in blocksize.
1242 Example:
1243 bsrange=1k-4k,2k-8k
1245 bssplit=str[,str][,str]
1247 Sometimes you want even finer grained control of the block sizes
1248 issued, not just an even split between them. This option allows
1249 you to weight various block sizes, so that you are able to de‐
1250 fine a specific amount of block sizes issued. The format for
1251 this option is:
1252 bssplit=blocksize/percentage:blocksize/percentage
1254 for as many block sizes as needed. So if you want to define a
1256 workload that has 50% 64k blocks, 10% 4k blocks, and 40% 32k
1257 blocks, you would write:
1258 bssplit=4k/10:64k/50:32k/40
1260 Ordering does not matter. If the percentage is left blank, fio
1262 will fill in the remaining values evenly. So a bssplit option
1263 like this one:
1264 bssplit=4k/50:1k/:32k/
1266 would have 50% 4k ios, and 25% 1k and 32k ios. The percentages
1268 always add up to 100, if bssplit is given a range that adds up
1269 to more, it will error out.
1270 Comma-separated values may be specified for reads, writes, and
1272 trims as described in blocksize.
1273 If you want a workload that has 50% 2k reads and 50% 4k reads,
1275 while having 90% 4k writes and 10% 8k writes, you would specify:
1276 bssplit=2k/50:4k/50,4k/90:8k/10
1278 Fio supports defining up to 64 different weights for each data
1280 direction.
1281 blocksize_unaligned, bs_unaligned
1283 If set, fio will issue I/O units with any size within block‐
1284 size_range, not just multiples of the minimum size. This typi‐
1285 cally won't work with direct I/O, as that normally requires sec‐
1286 tor alignment.
1287 bs_is_seq_rand=bool
1289 If this option is set, fio will use the normal read,write block‐
1290 size settings as sequential,random blocksize settings instead.
1291 Any random read or write will use the WRITE blocksize settings,
1292 and any sequential read or write will use the READ blocksize
1293 settings.
1294 blockalign=int[,int][,int], ba=int[,int][,int]
1296 Boundary to which fio will align random I/O units. Default:
1297 blocksize. Minimum alignment is typically 512b for using direct
1298 I/O, though it usually depends on the hardware block size. This
1299 option is mutually exclusive with using a random map for files,
1300 so it will turn off that option. Comma-separated values may be
1301 specified for reads, writes, and trims as described in block‐
1302 size.
1303 Buffers and memory
1305 zero_buffers
1306 Initialize buffers with all zeros. Default: fill buffers with
1307 random data.
1308 refill_buffers
1310 If this option is given, fio will refill the I/O buffers on ev‐
1311 ery submit. The default is to only fill it at init time and re‐
1312 use that data. Only makes sense if zero_buffers isn't specified,
1313 naturally. If data verification is enabled, refill_buffers is
1314 also automatically enabled.
1315 scramble_buffers=bool
1317 If refill_buffers is too costly and the target is using data
1318 deduplication, then setting this option will slightly modify the
1319 I/O buffer contents to defeat normal de-dupe attempts. This is
1320 not enough to defeat more clever block compression attempts, but
1321 it will stop naive dedupe of blocks. Default: true.
1322 buffer_compress_percentage=int
1324 If this is set, then fio will attempt to provide I/O buffer con‐
1325 tent (on WRITEs) that compresses to the specified level. Fio
1326 does this by providing a mix of random data followed by fixed
1327 pattern data. The fixed pattern is either zeros, or the pattern
1328 specified by buffer_pattern. If the buffer_pattern option is
1329 used, it might skew the compression ratio slightly. Setting buf‐
1330 fer_compress_percentage to a value other than 100 will also en‐
1331 able refill_buffers in order to reduce the likelihood that adja‐
1332 cent blocks are so similar that they over compress when seen to‐
1333 gether. See buffer_compress_chunk for how to set a finer or
1334 coarser granularity of the random/fixed data regions. Defaults
1335 to unset i.e., buffer data will not adhere to any compression
1336 level.
1337 buffer_compress_chunk=int
1339 This setting allows fio to manage how big the random/fixed data
1340 region is when using buffer_compress_percentage. When buf‐
1341 fer_compress_chunk is set to some non-zero value smaller than
1342 the block size, fio can repeat the random/fixed region through‐
1343 out the I/O buffer at the specified interval (which particularly
1344 useful when bigger block sizes are used for a job). When set to
1345 0, fio will use a chunk size that matches the block size result‐
1346 ing in a single random/fixed region within the I/O buffer. De‐
1347 faults to 512. When the unit is omitted, the value is inter‐
1348 preted in bytes.
1349 buffer_pattern=str
1351 If set, fio will fill the I/O buffers with this pattern or with
1352 the contents of a file. If not set, the contents of I/O buffers
1353 are defined by the other options related to buffer contents. The
1354 setting can be any pattern of bytes, and can be prefixed with 0x
1355 for hex values. It may also be a string, where the string must
1356 then be wrapped with "". Or it may also be a filename, where the
1357 filename must be wrapped with '' in which case the file is
1358 opened and read. Note that not all the file contents will be
1359 read if that would cause the buffers to overflow. So, for exam‐
1360 ple:
1361 buffer_pattern='filename'
1363 or:
1364 buffer_pattern="abcd"
1365 or:
1366 buffer_pattern=-12
1367 or:
1368 buffer_pattern=0xdeadface
1369 Also you can combine everything together in any order:
1371 buffer_pattern=0xdeadface"abcd"-12'filename'
1373 dedupe_percentage=int
1375 If set, fio will generate this percentage of identical buffers
1376 when writing. These buffers will be naturally dedupable. The
1377 contents of the buffers depend on what other buffer compression
1378 settings have been set. It's possible to have the individual
1379 buffers either fully compressible, or not at all -- this option
1380 only controls the distribution of unique buffers. Setting this
1381 option will also enable refill_buffers to prevent every buffer
1382 being identical.
1383 dedupe_mode=str
1385 If dedupe_percentage is given, then this option controls how fio
1386 generates the dedupe buffers.
1387 repeat
1389 Generate dedupe buffers by repeating previous
1391 writes
1392 working_set
1394 Generate dedupe buffers from working set
1396 repeat is the default option for fio. Dedupe buffers are gener‐
1398 ated by repeating previous unique write.
1399 working_set is a more realistic workload. With working_set,
1401 dedupe_working_set_percentage should be provided. Given that,
1402 fio will use the initial unique write buffers as its working
1403 set. Upon deciding to dedupe, fio will randomly choose a buffer
1404 from the working set. Note that by using working_set the dedupe
1405 percentage will converge to the desired over time while repeat
1406 maintains the desired percentage throughout the job.
1407 dedupe_working_set_percentage=int
1409 If dedupe_mode is set to working_set, then this controls the
1410 percentage of size of the file or device used as the buffers fio
1411 will choose to generate the dedupe buffers from
1412 Note that size needs to be explicitly provided and only 1 file
1414 per job is supported
1415 invalidate=bool
1417 Invalidate the buffer/page cache parts of the files to be used
1418 prior to starting I/O if the platform and file type support it.
1419 Defaults to true. This will be ignored if pre_read is also
1420 specified for the same job.
1421 sync=str
1423 Whether, and what type, of synchronous I/O to use for writes.
1424 The allowed values are:
1425 none Do not use synchronous IO, the default.
1427 0 Same as none.
1429 sync Use synchronous file IO. For the majority of I/O
1431 engines, this means using O_SYNC.
1432 1 Same as sync.
1434 dsync Use synchronous data IO. For the majority of I/O
1436 engines, this means using O_DSYNC.
1437 iomem=str, mem=str
1439 Fio can use various types of memory as the I/O unit buffer. The
1440 allowed values are:
1441 malloc Use memory from malloc(3) as the buffers. Default
1443 memory type.
1444 shm Use shared memory as the buffers. Allocated
1446 through shmget(2).
1447 shmhuge
1449 Same as shm, but use huge pages as backing.
1450 mmap Use mmap(2) to allocate buffers. May either be
1452 anonymous memory, or can be file backed if a file‐
1453 name is given after the option. The format is
1454 `mem=mmap:/path/to/file'.
1455 mmaphuge
1457 Use a memory mapped huge file as the buffer back‐
1458 ing. Append filename after mmaphuge, ala `mem=mma‐
1459 phuge:/hugetlbfs/file'.
1460 mmapshared
1462 Same as mmap, but use a MMAP_SHARED mapping.
1463 cudamalloc
1465 Use GPU memory as the buffers for GPUDirect RDMA
1466 benchmark. The ioengine must be rdma.
1467 The area allocated is a function of the maximum allowed bs size
1469 for the job, multiplied by the I/O depth given. Note that for
1470 shmhuge and mmaphuge to work, the system must have free huge
1471 pages allocated. This can normally be checked and set by read‐
1472 ing/writing `/proc/sys/vm/nr_hugepages' on a Linux system. Fio
1473 assumes a huge page is 4MiB in size. So to calculate the number
1474 of huge pages you need for a given job file, add up the I/O
1475 depth of all jobs (normally one unless iodepth is used) and mul‐
1476 tiply by the maximum bs set. Then divide that number by the huge
1477 page size. You can see the size of the huge pages in `/proc/mem‐
1478 info'. If no huge pages are allocated by having a non-zero num‐
1479 ber in `nr_hugepages', using mmaphuge or shmhuge will fail. Also
1480 see hugepage-size.
1481 mmaphuge also needs to have hugetlbfs mounted and the file loca‐
1483 tion should point there. So if it's mounted in `/huge', you
1484 would use `mem=mmaphuge:/huge/somefile'.
1485 iomem_align=int, mem_align=int
1487 This indicates the memory alignment of the I/O memory buffers.
1488 Note that the given alignment is applied to the first I/O unit
1489 buffer, if using iodepth the alignment of the following buffers
1490 are given by the bs used. In other words, if using a bs that is
1491 a multiple of the page sized in the system, all buffers will be
1492 aligned to this value. If using a bs that is not page aligned,
1493 the alignment of subsequent I/O memory buffers is the sum of the
1494 iomem_align and bs used.
1495 hugepage-size=int
1497 Defines the size of a huge page. Must at least be equal to the
1498 system setting, see `/proc/meminfo'. Defaults to 4MiB. Should
1499 probably always be a multiple of megabytes, so using
1500 `hugepage-size=Xm' is the preferred way to set this to avoid
1501 setting a non-pow-2 bad value.
1502 lockmem=int
1504 Pin the specified amount of memory with mlock(2). Can be used to
1505 simulate a smaller amount of memory. The amount specified is per
1506 worker.
1507 I/O size
1509 size=int[%|z]
1510 The total size of file I/O for each thread of this job. Fio will
1511 run until this many bytes has been transferred, unless runtime
1512 is limited by other options (such as runtime, for instance, or
1513 increased/decreased by io_size). Fio will divide this size be‐
1514 tween the available files determined by options such as nrfiles,
1515 filename, unless filesize is specified by the job. If the result
1516 of division happens to be 0, the size is set to the physical
1517 size of the given files or devices if they exist. If this op‐
1518 tion is not specified, fio will use the full size of the given
1519 files or devices. If the files do not exist, size must be given.
1520 It is also possible to give size as a percentage between 1 and
1521 100. If `size=20%' is given, fio will use 20% of the full size
1522 of the given files or devices. In ZBD mode, size can be given in
1523 units of number of zones using 'z'. Can be combined with offset
1524 to constrain the start and end range that I/O will be done
1525 within.
1526 io_size=int[%|z], io_limit=int[%|z]
1528 Normally fio operates within the region set by size, which means
1529 that the size option sets both the region and size of I/O to be
1530 performed. Sometimes that is not what you want. With this op‐
1531 tion, it is possible to define just the amount of I/O that fio
1532 should do. For instance, if size is set to 20GiB and io_size is
1533 set to 5GiB, fio will perform I/O within the first 20GiB but
1534 exit when 5GiB have been done. The opposite is also possible --
1535 if size is set to 20GiB, and io_size is set to 40GiB, then fio
1536 will do 40GiB of I/O within the 0..20GiB region. Value can be
1537 set as percentage: io_size=N%. In this case io_size multiplies
1538 size= value. In ZBD mode, value can also be set as number of
1539 zones using 'z'.
1540 filesize=irange(int)
1542 Individual file sizes. May be a range, in which case fio will
1543 select sizes for files at random within the given range and lim‐
1544 ited to size in total (if that is given). If not given, each
1545 created file is the same size. This option overrides size in
1546 terms of file size, which means this value is used as a fixed
1547 size or possible range of each file.
1548 file_append=bool
1550 Perform I/O after the end of the file. Normally fio will operate
1551 within the size of a file. If this option is set, then fio will
1552 append to the file instead. This has identical behavior to set‐
1553 ting offset to the size of a file. This option is ignored on
1554 non-regular files.
1555 fill_device=bool, fill_fs=bool
1557 Sets size to something really large and waits for ENOSPC (no
1558 space left on device) or EDQUOT (disk quota exceeded) as the
1559 terminating condition. Only makes sense with sequential write.
1560 For a read workload, the mount point will be filled first then
1561 I/O started on the result.
1562 I/O engine
1564 ioengine=str
1565 Defines how the job issues I/O to the file. The following types
1566 are defined:
1567 sync Basic read(2) or write(2) I/O. lseek(2) is used to
1569 position the I/O location. See fsync and fdata‐
1570 sync for syncing write I/Os.
1571 psync Basic pread(2) or pwrite(2) I/O. Default on all
1573 supported operating systems except for Windows.
1574 vsync Basic readv(2) or writev(2) I/O. Will emulate
1576 queuing by coalescing adjacent I/Os into a single
1577 submission.
1578 pvsync Basic preadv(2) or pwritev(2) I/O.
1580 pvsync2
1582 Basic preadv2(2) or pwritev2(2) I/O.
1583 libaio Linux native asynchronous I/O. Note that Linux may
1585 only support queued behavior with non-buffered I/O
1586 (set `direct=1' or `buffered=0'). This engine de‐
1587 fines engine specific options.
1588 posixaio
1590 POSIX asynchronous I/O using aio_read(3) and
1591 aio_write(3).
1592 solarisaio
1594 Solaris native asynchronous I/O.
1595 windowsaio
1597 Windows native asynchronous I/O. Default on Win‐
1598 dows.
1599 mmap File is memory mapped with mmap(2) and data copied
1601 to/from using memcpy(3).
1602 splice splice(2) is used to transfer the data and vm‐
1604 splice(2) to transfer data from user space to the
1605 kernel.
1606 sg SCSI generic sg v3 I/O. May either be synchronous
1608 using the SG_IO ioctl, or if the target is an sg
1609 character device we use read(2) and write(2) for
1610 asynchronous I/O. Requires filename option to
1611 specify either block or character devices. This
1612 engine supports trim operations. The sg engine in‐
1613 cludes engine specific options.
1614 libzbc Read, write, trim and ZBC/ZAC operations to a
1616 zoned block device using libzbc library. The tar‐
1617 get can be either an SG character device or a
1618 block device file.
1619 null Doesn't transfer any data, just pretends to. This
1621 is mainly used to exercise fio itself and for de‐
1622 bugging/testing purposes.
1623 net Transfer over the network to given `host:port'.
1625 Depending on the protocol used, the hostname,
1626 port, listen and filename options are used to
1627 specify what sort of connection to make, while the
1628 protocol option determines which protocol will be
1629 used. This engine defines engine specific options.
1630 netsplice
1632 Like net, but uses splice(2) and vmsplice(2) to
1633 map data and send/receive. This engine defines
1634 engine specific options.
1635 cpuio Doesn't transfer any data, but burns CPU cycles
1637 according to the cpuload, cpuchunks and cpumode
1638 options. A job never finishes unless there is at
1639 least one non-cpuio job.
1640 cpuload=85 will cause that job to do nothing but
1642 burn 85% of the CPU. In case of SMP machines, use
1643 numjobs=<nr_of_cpu> to get desired CPU usage, as
1644 the cpuload only loads a single CPU at the desired
1645 rate.
1646 cpumode=qsort replace the default noop instruc‐
1648 tions loop by a qsort algorithm to consume more
1649 energy.
1650 rdma The RDMA I/O engine supports both RDMA memory se‐
1652 mantics (RDMA_WRITE/RDMA_READ) and channel seman‐
1653 tics (Send/Recv) for the InfiniBand, RoCE and
1654 iWARP protocols. This engine defines engine spe‐
1655 cific options.
1656 falloc I/O engine that does regular fallocate to simulate
1657 data transfer as fio ioengine.
1658 DDIR_READ does fallocate(,mode = FAL‐
1659 LOC_FL_KEEP_SIZE,).
1660 DIR_WRITE does fallocate(,mode = 0).
1661 DDIR_TRIM does fallocate(,mode = FAL‐
1662 LOC_FL_KEEP_SIZE|FALLOC_FL_PUNCH_HOLE).
1663 ftruncate
1665 I/O engine that sends ftruncate(2) operations in
1666 response to write (DDIR_WRITE) events. Each ftrun‐
1667 cate issued sets the file's size to the current
1668 block offset. blocksize is ignored.
1669 e4defrag
1671 I/O engine that does regular EXT4_IOC_MOVE_EXT
1672 ioctls to simulate defragment activity in request
1673 to DDIR_WRITE event.
1674 rados I/O engine supporting direct access to Ceph Reli‐
1676 able Autonomic Distributed Object Store (RADOS)
1677 via librados. This ioengine defines engine spe‐
1678 cific options.
1679 rbd I/O engine supporting direct access to Ceph Rados
1681 Block Devices (RBD) via librbd without the need to
1682 use the kernel rbd driver. This ioengine defines
1683 engine specific options.
1684 http I/O engine supporting GET/PUT requests over
1686 HTTP(S) with libcurl to a WebDAV or S3 endpoint.
1687 This ioengine defines engine specific options.
1688 This engine only supports direct IO of iodepth=1;
1690 you need to scale this via numjobs. blocksize de‐
1691 fines the size of the objects to be created.
1692 TRIM is translated to object deletion.
1694 gfapi Using GlusterFS libgfapi sync interface to direct
1696 access to GlusterFS volumes without having to go
1697 through FUSE. This ioengine defines engine spe‐
1698 cific options.
1699 gfapi_async
1701 Using GlusterFS libgfapi async interface to direct
1702 access to GlusterFS volumes without having to go
1703 through FUSE. This ioengine defines engine spe‐
1704 cific options.
1705 libhdfs
1707 Read and write through Hadoop (HDFS). The filename
1708 option is used to specify host,port of the hdfs
1709 name-node to connect. This engine interprets off‐
1710 sets a little differently. In HDFS, files once
1711 created cannot be modified so random writes are
1712 not possible. To imitate this the libhdfs engine
1713 expects a bunch of small files to be created over
1714 HDFS and will randomly pick a file from them based
1715 on the offset generated by fio backend (see the
1716 example job file to create such files, use
1717 `rw=write' option). Please note, it may be neces‐
1718 sary to set environment variables to work with
1719 HDFS/libhdfs properly. Each job uses its own con‐
1720 nection to HDFS.
1721 mtd Read, write and erase an MTD character device
1723 (e.g., `/dev/mtd0'). Discards are treated as
1724 erases. Depending on the underlying device type,
1725 the I/O may have to go in a certain pattern, e.g.,
1726 on NAND, writing sequentially to erase blocks and
1727 discarding before overwriting. The trimwrite mode
1728 works well for this constraint.
1729 pmemblk
1731 Read and write using filesystem DAX to a file on a
1732 filesystem mounted with DAX on a persistent memory
1733 device through the PMDK libpmemblk library.
1734 dev-dax
1736 Read and write using device DAX to a persistent
1737 memory device (e.g., /dev/dax0.0) through the PMDK
1738 libpmem library.
1739 external
1741 Prefix to specify loading an external I/O engine
1742 object file. Append the engine filename, e.g. `io‐
1743 engine=external:/tmp/foo.o' to load ioengine
1744 `foo.o' in `/tmp'. The path can be either absolute
1745 or relative. See `engines/skeleton_external.c' in
1746 the fio source for details of writing an external
1747 I/O engine.
1748 filecreate
1750 Simply create the files and do no I/O to them.
1751 You still need to set filesize so that all the ac‐
1752 counting still occurs, but no actual I/O will be
1753 done other than creating the file.
1754 filestat
1756 Simply do stat() and do no I/O to the file. You
1757 need to set 'filesize' and 'nrfiles', so that
1758 files will be created. This engine is to measure
1759 file lookup and meta data access.
1760 filedelete
1762 Simply delete files by unlink() and do no I/O to
1763 the file. You need to set 'filesize' and 'nr‐
1764 files', so that files will be created. This en‐
1765 gine is to measure file delete.
1766 libpmem
1768 Read and write using mmap I/O to a file on a
1769 filesystem mounted with DAX on a persistent memory
1770 device through the PMDK libpmem library.
1771 ime_psync
1773 Synchronous read and write using DDN's Infinite
1774 Memory Engine (IME). This engine is very basic and
1775 issues calls to IME whenever an IO is queued.
1776 ime_psyncv
1778 Synchronous read and write using DDN's Infinite
1779 Memory Engine (IME). This engine uses iovecs and
1780 will try to stack as much IOs as possible (if the
1781 IOs are "contiguous" and the IO depth is not ex‐
1782 ceeded) before issuing a call to IME.
1783 ime_aio
1785 Asynchronous read and write using DDN's Infinite
1786 Memory Engine (IME). This engine will try to stack
1787 as much IOs as possible by creating requests for
1788 IME. FIO will then decide when to commit these
1789 requests.
1790 libiscsi
1792 Read and write iscsi lun with libiscsi.
1793 nbd Synchronous read and write a Network Block Device
1795 (NBD).
1796 libcufile
1798 I/O engine supporting libcufile synchronous access
1799 to nvidia-fs and a GPUDirect Storage-supported
1800 filesystem. This engine performs I/O without
1801 transferring buffers between user-space and the
1802 kernel, unless verify is set or cuda_io is posix.
1803 iomem must not be cudamalloc. This ioengine de‐
1804 fines engine specific options.
1805 dfs I/O engine supporting asynchronous read and write
1807 operations to the DAOS File System (DFS) via
1808 libdfs.
1809 nfs I/O engine supporting asynchronous read and write
1811 operations to NFS filesystems from userspace via
1812 libnfs. This is useful for achieving higher con‐
1813 currency and thus throughput than is possible via
1814 kernel NFS.
1815 exec Execute 3rd party tools. Could be used to perform
1817 monitoring during jobs runtime.
1818 I/O engine specific parameters
1820 In addition, there are some parameters which are only valid when a spe‐
1821 cific ioengine is in use. These are used identically to normal parame‐
1822 ters, with the caveat that when used on the command line, they must
1823 come after the ioengine that defines them is selected.
1824 (io_uring,libaio)cmdprio_percentage=int[,int]
1826 Set the percentage of I/O that will be issued with the highest
1827 priority. Default: 0. A single value applies to reads and
1828 writes. Comma-separated values may be specified for reads and
1829 writes. For this option to be effective, NCQ priority must be
1830 supported and enabled, and `direct=1' option must be used. fio
1831 must also be run as the root user. Unlike slat/clat/lat stats,
1832 which can be tracked and reported independently, per priority
1833 stats only track and report a single type of latency. By de‐
1834 fault, completion latency (clat) will be reported, if lat_per‐
1835 centiles is set, total latency (lat) will be reported.
1836 (io_uring,libaio)cmdprio_class=int[,int]
1838 Set the I/O priority class to use for I/Os that must be issued
1839 with a priority when cmdprio_percentage or cmdprio_bssplit is
1840 set. If not specified when cmdprio_percentage or cmdprio_bss‐
1841 plit is set, this defaults to the highest priority class. A sin‐
1842 gle value applies to reads and writes. Comma-separated values
1843 may be specified for reads and writes. See man ionice(1). See
1844 also the prioclass option.
1845 (io_uring,libaio)cmdprio=int[,int]
1847 Set the I/O priority value to use for I/Os that must be issued
1848 with a priority when cmdprio_percentage or cmdprio_bssplit is
1849 set. If not specified when cmdprio_percentage or cmdprio_bss‐
1850 plit is set, this defaults to 0. Linux limits us to a positive
1851 value between 0 and 7, with 0 being the highest. A single value
1852 applies to reads and writes. Comma-separated values may be
1853 specified for reads and writes. See man ionice(1). Refer to an
1854 appropriate manpage for other operating systems since the mean‐
1855 ing of priority may differ. See also the prio option.
1856 (io_uring,libaio)cmdprio_bssplit=str[,str]
1858 To get a finer control over I/O priority, this option allows
1859 specifying the percentage of IOs that must have a priority set
1860 depending on the block size of the IO. This option is useful
1861 only when used together with the option bssplit, that is, multi‐
1862 ple different block sizes are used for reads and writes. The
1863 format for this option is the same as the format of the bssplit
1864 option, with the exception that values for trim IOs are ignored.
1865 This option is mutually exclusive with the cmdprio_percentage
1866 option.
1867 (io_uring)fixedbufs
1869 If fio is asked to do direct IO, then Linux will map pages for
1870 each IO call, and release them when IO is done. If this option
1871 is set, the pages are pre-mapped before IO is started. This
1872 eliminates the need to map and release for each IO. This is
1873 more efficient, and reduces the IO latency as well.
1874 (io_uring)hipri
1876 If this option is set, fio will attempt to use polled IO comple‐
1877 tions. Normal IO completions generate interrupts to signal the
1878 completion of IO, polled completions do not. Hence they are re‐
1879 quire active reaping by the application. The benefits are more
1880 efficient IO for high IOPS scenarios, and lower latencies for
1881 low queue depth IO.
1882 (io_uring)registerfiles
1884 With this option, fio registers the set of files being used with
1885 the kernel. This avoids the overhead of managing file counts in
1886 the kernel, making the submission and completion part more
1887 lightweight. Required for the below sqthread_poll option.
1888 (io_uring)sqthread_poll
1890 Normally fio will submit IO by issuing a system call to notify
1891 the kernel of available items in the SQ ring. If this option is
1892 set, the act of submitting IO will be done by a polling thread
1893 in the kernel. This frees up cycles for fio, at the cost of us‐
1894 ing more CPU in the system.
1895 (io_uring)sqthread_poll_cpu
1897 When `sqthread_poll` is set, this option provides a way to de‐
1898 fine which CPU should be used for the polling thread.
1899 (libaio)userspace_reap
1901 Normally, with the libaio engine in use, fio will use the
1902 io_getevents(3) system call to reap newly returned events. With
1903 this flag turned on, the AIO ring will be read directly from
1904 user-space to reap events. The reaping mode is only enabled when
1905 polling for a minimum of 0 events (e.g. when `iodepth_batch_com‐
1906 plete=0').
1907 (pvsync2)hipri
1909 Set RWF_HIPRI on I/O, indicating to the kernel that it's of
1910 higher priority than normal.
1911 (pvsync2)hipri_percentage
1913 When hipri is set this determines the probability of a pvsync2
1914 I/O being high priority. The default is 100%.
1915 (pvsync2,libaio,io_uring)nowait
1917 By default if a request cannot be executed immediately (e.g. re‐
1918 source starvation, waiting on locks) it is queued and the initi‐
1919 ating process will be blocked until the required resource be‐
1920 comes free. This option sets the RWF_NOWAIT flag (supported
1921 from the 4.14 Linux kernel) and the call will return instantly
1922 with EAGAIN or a partial result rather than waiting.
1923 It is useful to also use ignore_error=EAGAIN when using this op‐
1925 tion. Note: glibc 2.27, 2.28 have a bug in syscall wrappers
1926 preadv2, pwritev2. They return EOPNOTSUP instead of EAGAIN.
1927 For cached I/O, using this option usually means a request oper‐
1929 ates only with cached data. Currently the RWF_NOWAIT flag does
1930 not supported for cached write. For direct I/O, requests will
1931 only succeed if cache invalidation isn't required, file blocks
1932 are fully allocated and the disk request could be issued immedi‐
1933 ately.
1934 (cpuio)cpuload=int
1936 Attempt to use the specified percentage of CPU cycles. This is a
1937 mandatory option when using cpuio I/O engine.
1938 (cpuio)cpuchunks=int
1940 Split the load into cycles of the given time. In microseconds.
1941 (cpuio)exit_on_io_done=bool
1943 Detect when I/O threads are done, then exit.
1944 (libhdfs)namenode=str
1946 The hostname or IP address of a HDFS cluster namenode to con‐
1947 tact.
1948 (libhdfs)port=int
1950 The listening port of the HFDS cluster namenode.
1951 (netsplice,net)port=int
1953 The TCP or UDP port to bind to or connect to. If this is used
1954 with numjobs to spawn multiple instances of the same job type,
1955 then this will be the starting port number since fio will use a
1956 range of ports.
1957 (rdma,librpma_*)port=int
1959 The port to use for RDMA-CM communication. This should be the
1960 same value on the client and the server side.
1961 (netsplice,net,rdma)hostname=str
1963 The hostname or IP address to use for TCP, UDP or RDMA-CM based
1964 I/O. If the job is a TCP listener or UDP reader, the hostname
1965 is not used and must be omitted unless it is a valid UDP multi‐
1966 cast address.
1967 (librpma_*)serverip=str
1969 The IP address to be used for RDMA-CM based I/O.
1970 (librpma_*_server)direct_write_to_pmem=bool
1972 Set to 1 only when Direct Write to PMem from the remote host is
1973 possible. Otherwise, set to 0.
1974 (librpma_*_server)busy_wait_polling=bool
1976 Set to 0 to wait for completion instead of busy-wait polling
1977 completion. Default: 1.
1978 (netsplice,net)interface=str
1980 The IP address of the network interface used to send or receive
1981 UDP multicast.
1982 (netsplice,net)ttl=int
1984 Time-to-live value for outgoing UDP multicast packets. Default:
1985 1.
1986 (netsplice,net)nodelay=bool
1988 Set TCP_NODELAY on TCP connections.
1989 (netsplice,net)protocol=str, proto=str
1991 The network protocol to use. Accepted values are:
1992 tcp Transmission control protocol.
1994 tcpv6 Transmission control protocol V6.
1996 udp User datagram protocol.
1998 udpv6 User datagram protocol V6.
2000 unix UNIX domain socket.
2002 When the protocol is TCP or UDP, the port must also be given, as
2004 well as the hostname if the job is a TCP listener or UDP reader.
2005 For unix sockets, the normal filename option should be used and
2006 the port is invalid.
2007 (netsplice,net)listen
2009 For TCP network connections, tell fio to listen for incoming
2010 connections rather than initiating an outgoing connection. The
2011 hostname must be omitted if this option is used.
2012 (netsplice,net)pingpong
2014 Normally a network writer will just continue writing data, and a
2015 network reader will just consume packages. If `pingpong=1' is
2016 set, a writer will send its normal payload to the reader, then
2017 wait for the reader to send the same payload back. This allows
2018 fio to measure network latencies. The submission and completion
2019 latencies then measure local time spent sending or receiving,
2020 and the completion latency measures how long it took for the
2021 other end to receive and send back. For UDP multicast traffic
2022 `pingpong=1' should only be set for a single reader when multi‐
2023 ple readers are listening to the same address.
2024 (netsplice,net)window_size=int
2026 Set the desired socket buffer size for the connection.
2027 (netsplice,net)mss=int
2029 Set the TCP maximum segment size (TCP_MAXSEG).
2030 (e4defrag)donorname=str
2032 File will be used as a block donor (swap extents between files).
2033 (e4defrag)inplace=int
2035 Configure donor file blocks allocation strategy:
2036 0 Default. Preallocate donor's file on init.
2038 1 Allocate space immediately inside defragment
2040 event, and free right after event.
2041 (rbd,rados)clustername=str
2043 Specifies the name of the Ceph cluster.
2044 (rbd)rbdname=str
2046 Specifies the name of the RBD.
2047 (rbd,rados)pool=str
2049 Specifies the name of the Ceph pool containing RBD or RADOS
2050 data.
2051 (rbd,rados)clientname=str
2053 Specifies the username (without the 'client.' prefix) used to
2054 access the Ceph cluster. If the clustername is specified, the
2055 clientname shall be the full *type.id* string. If no type. pre‐
2056 fix is given, fio will add 'client.' by default.
2057 (rbd,rados)busy_poll=bool
2059 Poll store instead of waiting for completion. Usually this pro‐
2060 vides better throughput at cost of higher(up to 100%) CPU uti‐
2061 lization.
2062 (rados)touch_objects=bool
2064 During initialization, touch (create if do not exist) all ob‐
2065 jects (files). Touching all objects affects ceph caches and
2066 likely impacts test results. Enabled by default.
2067 (http)http_host=str
2069 Hostname to connect to. For S3, this could be the bucket name.
2070 Default is localhost
2071 (http)http_user=str
2073 Username for HTTP authentication.
2074 (http)http_pass=str
2076 Password for HTTP authentication.
2077 (http)https=str
2079 Whether to use HTTPS instead of plain HTTP. on enables HTTPS;
2080 insecure will enable HTTPS, but disable SSL peer verification
2081 (use with caution!). Default is off.
2082 (http)http_mode=str
2084 Which HTTP access mode to use: webdav, swift, or s3. Default is
2085 webdav.
2086 (http)http_s3_region=str
2088 The S3 region/zone to include in the request. Default is us-
2089 east-1.
2090 (http)http_s3_key=str
2092 The S3 secret key.
2093 (http)http_s3_keyid=str
2095 The S3 key/access id.
2096 (http)http_swift_auth_token=str
2098 The Swift auth token. See the example configuration file on how
2099 to retrieve this.
2100 (http)http_verbose=int
2102 Enable verbose requests from libcurl. Useful for debugging. 1
2103 turns on verbose logging from libcurl, 2 additionally enables
2104 HTTP IO tracing. Default is 0
2105 (mtd)skip_bad=bool
2107 Skip operations against known bad blocks.
2108 (libhdfs)hdfsdirectory
2110 libhdfs will create chunk in this HDFS directory.
2111 (libhdfs)chunk_size
2113 The size of the chunk to use for each file.
2114 (rdma)verb=str
2116 The RDMA verb to use on this side of the RDMA ioengine connec‐
2117 tion. Valid values are write, read, send and recv. These corre‐
2118 spond to the equivalent RDMA verbs (e.g. write = rdma_write
2119 etc.). Note that this only needs to be specified on the client
2120 side of the connection. See the examples folder.
2121 (rdma)bindname=str
2123 The name to use to bind the local RDMA-CM connection to a local
2124 RDMA device. This could be a hostname or an IPv4 or IPv6 ad‐
2125 dress. On the server side this will be passed into the
2126 rdma_bind_addr() function and on the client site it will be used
2127 in the rdma_resolve_add() function. This can be useful when mul‐
2128 tiple paths exist between the client and the server or in cer‐
2129 tain loopback configurations.
2130 (filestat)stat_type=str
2132 Specify stat system call type to measure lookup/getattr perfor‐
2133 mance. Default is stat for stat(2).
2134 (sg)hipri
2136 If this option is set, fio will attempt to use polled IO comple‐
2137 tions. This will have a similar effect as (io_uring)hipri. Only
2138 SCSI READ and WRITE commands will have the SGV4_FLAG_HIPRI set
2139 (not UNMAP (trim) nor VERIFY). Older versions of the Linux sg
2140 driver that do not support hipri will simply ignore this flag
2141 and do normal IO. The Linux SCSI Low Level Driver (LLD) that
2142 "owns" the device also needs to support hipri (also known as
2143 iopoll and mq_poll). The MegaRAID driver is an example of a SCSI
2144 LLD. Default: clear (0) which does normal (interrupted based)
2145 IO.
2146 (sg)readfua=bool
2148 With readfua option set to 1, read operations include the force
2149 unit access (fua) flag. Default: 0.
2150 (sg)writefua=bool
2152 With writefua option set to 1, write operations include the
2153 force unit access (fua) flag. Default: 0.
2154 (sg)sg_write_mode=str
2156 Specify the type of write commands to issue. This option can
2157 take three values:
2158 write (default)
2160 Write opcodes are issued as usual
2161 verify Issue WRITE AND VERIFY commands. The BYTCHK bit is
2163 set to 0. This directs the device to carry out a
2164 medium verification with no data comparison. The
2165 writefua option is ignored with this selection.
2166 same Issue WRITE SAME commands. This transfers a single
2168 block to the device and writes this same block of
2169 data to a contiguous sequence of LBAs beginning at
2170 the specified offset. fio's block size parameter
2171 specifies the amount of data written with each
2172 command. However, the amount of data actually
2173 transferred to the device is equal to the device's
2174 block (sector) size. For a device with 512 byte
2175 sectors, blocksize=8k will write 16 sectors with
2176 each command. fio will still generate 8k of data
2177 for each command butonly the first 512 bytes will
2178 be used and transferred to the device. The write‐
2179 fua option is ignored with this selection.
2180 (nbd)uri=str
2182 Specify the NBD URI of the server to test. The string is a
2183 standard NBD URI (see https://github.com/NetworkBlockDe‐
2184 vice/nbd/tree/master/doc). Example URIs:
2185 nbd://localhost:10809
2187 nbd+unix:///?socket=/tmp/socket
2189 nbds://tlshost/exportname
2191 (libcufile)gpu_dev_ids=str
2193 Specify the GPU IDs to use with CUDA. This is a colon-separated
2194 list of int. GPUs are assigned to workers roundrobin. Default
2195 is 0.
2196 (libcufile)cuda_io=str
2198 Specify the type of I/O to use with CUDA. This option takes the
2199 following values:
2200 cufile (default)
2202 Use libcufile and nvidia-fs. This option performs
2203 I/O directly between a GPUDirect Storage filesys‐
2204 tem and GPU buffers, avoiding use of a bounce buf‐
2205 fer. If verify is set, cudaMemcpy is used to copy
2206 verification data between RAM and GPU(s). Verifi‐
2207 cation data is copied from RAM to GPU before a
2208 write and from GPU to RAM after a read. direct
2209 must be 1.
2210 posix Use POSIX to perform I/O with a RAM buffer, and
2212 use cudaMemcpy to transfer data between RAM and
2213 the GPU(s). Data is copied from GPU to RAM before
2214 a write and copied from RAM to GPU after a read.
2215 verify does not affect the use of cudaMemcpy.
2216 (dfs)pool
2218 Specify the label or UUID of the DAOS pool to connect to.
2219 (dfs)cont
2221 Specify the label or UUID of the DAOS container to open.
2222 (dfs)chunk_size
2224 Specificy a different chunk size (in bytes) for the dfs file.
2225 Use DAOS container's chunk size by default.
2226 (dfs)object_class
2228 Specificy a different object class for the dfs file. Use DAOS
2229 container's object class by default.
2230 (nfs)nfs_url
2232 URL in libnfs format, eg
2233 nfs://<server|ipv4|ipv6>/path[?arg=val[&arg=val]*] Refer to the
2234 libnfs README for more details.
2235 (exec)program=str
2237 Specify the program to execute. Note the program will receive a
2238 SIGTERM when the job is reaching the time limit. A SIGKILL is
2239 sent once the job is over. The delay between the two signals is
2240 defined by grace_time option.
2241 (exec)arguments=str
2243 Specify arguments to pass to program. Some special variables
2244 can be expanded to pass fio's job details to the program :
2245 %r replaced by the duration of the job in seconds
2247 %n replaced by the name of the job
2249 (exec)grace_time=int
2251 Defines the time between the SIGTERM and SIGKILL signals. De‐
2252 fault is 1 second.
2253 (exec)std_redirect=ool
2255 If set, stdout and stderr streams are redirected to files named
2256 from the job name. Default is true.
2257 I/O depth
2259 iodepth=int
2260 Number of I/O units to keep in flight against the file. Note
2261 that increasing iodepth beyond 1 will not affect synchronous io‐
2262 engines (except for small degrees when verify_async is in use).
2263 Even async engines may impose OS restrictions causing the de‐
2264 sired depth not to be achieved. This may happen on Linux when
2265 using libaio and not setting `direct=1', since buffered I/O is
2266 not async on that OS. Keep an eye on the I/O depth distribution
2267 in the fio output to verify that the achieved depth is as ex‐
2268 pected. Default: 1.
2269 iodepth_batch_submit=int, iodepth_batch=int
2271 This defines how many pieces of I/O to submit at once. It de‐
2272 faults to 1 which means that we submit each I/O as soon as it is
2273 available, but can be raised to submit bigger batches of I/O at
2274 the time. If it is set to 0 the iodepth value will be used.
2275 iodepth_batch_complete_min=int, iodepth_batch_complete=int
2277 This defines how many pieces of I/O to retrieve at once. It de‐
2278 faults to 1 which means that we'll ask for a minimum of 1 I/O in
2279 the retrieval process from the kernel. The I/O retrieval will go
2280 on until we hit the limit set by iodepth_low. If this variable
2281 is set to 0, then fio will always check for completed events be‐
2282 fore queuing more I/O. This helps reduce I/O latency, at the
2283 cost of more retrieval system calls.
2284 iodepth_batch_complete_max=int
2286 This defines maximum pieces of I/O to retrieve at once. This
2287 variable should be used along with iodepth_batch_com‐
2288 plete_min=int variable, specifying the range of min and max
2289 amount of I/O which should be retrieved. By default it is equal
2290 to iodepth_batch_complete_min value. Example #1:
2291 iodepth_batch_complete_min=1
2293 iodepth_batch_complete_max=<iodepth>
2294 which means that we will retrieve at least 1 I/O and up to the
2296 whole submitted queue depth. If none of I/O has been completed
2297 yet, we will wait. Example #2:
2298 iodepth_batch_complete_min=0
2300 iodepth_batch_complete_max=<iodepth>
2301 which means that we can retrieve up to the whole submitted queue
2303 depth, but if none of I/O has been completed yet, we will NOT
2304 wait and immediately exit the system call. In this example we
2305 simply do polling.
2306 iodepth_low=int
2308 The low water mark indicating when to start filling the queue
2309 again. Defaults to the same as iodepth, meaning that fio will
2310 attempt to keep the queue full at all times. If iodepth is set
2311 to e.g. 16 and iodepth_low is set to 4, then after fio has
2312 filled the queue of 16 requests, it will let the depth drain
2313 down to 4 before starting to fill it again.
2314 serialize_overlap=bool
2316 Serialize in-flight I/Os that might otherwise cause or suffer
2317 from data races. When two or more I/Os are submitted simultane‐
2318 ously, there is no guarantee that the I/Os will be processed or
2319 completed in the submitted order. Further, if two or more of
2320 those I/Os are writes, any overlapping region between them can
2321 become indeterminate/undefined on certain storage. These issues
2322 can cause verification to fail erratically when at least one of
2323 the racing I/Os is changing data and the overlapping region has
2324 a non-zero size. Setting serialize_overlap tells fio to avoid
2325 provoking this behavior by explicitly serializing in-flight I/Os
2326 that have a non-zero overlap. Note that setting this option can
2327 reduce both performance and the iodepth achieved.
2328 This option only applies to I/Os issued for a single job except
2330 when it is enabled along with io_submit_mode=offload. In offload
2331 mode, fio will check for overlap among all I/Os submitted by
2332 offload jobs with serialize_overlap enabled.
2333 Default: false.
2335 io_submit_mode=str
2337 This option controls how fio submits the I/O to the I/O engine.
2338 The default is `inline', which means that the fio job threads
2339 submit and reap I/O directly. If set to `offload', the job
2340 threads will offload I/O submission to a dedicated pool of I/O
2341 threads. This requires some coordination and thus has a bit of
2342 extra overhead, especially for lower queue depth I/O where it
2343 can increase latencies. The benefit is that fio can manage sub‐
2344 mission rates independently of the device completion rates. This
2345 avoids skewed latency reporting if I/O gets backed up on the de‐
2346 vice side (the coordinated omission problem). Note that this op‐
2347 tion cannot reliably be used with async IO engines.
2348 I/O rate
2350 thinktime=time
2351 Stall the job for the specified period of time after an I/O has
2352 completed before issuing the next. May be used to simulate pro‐
2353 cessing being done by an application. When the unit is omitted,
2354 the value is interpreted in microseconds. See thinktime_blocks,
2355 thinktime_iotime and thinktime_spin.
2356 thinktime_spin=time
2358 Only valid if thinktime is set - pretend to spend CPU time doing
2359 something with the data received, before falling back to sleep‐
2360 ing for the rest of the period specified by thinktime. When the
2361 unit is omitted, the value is interpreted in microseconds.
2362 thinktime_blocks=int
2364 Only valid if thinktime is set - control how many blocks to is‐
2365 sue, before waiting thinktime usecs. If not set, defaults to 1
2366 which will make fio wait thinktime usecs after every block. This
2367 effectively makes any queue depth setting redundant, since no
2368 more than 1 I/O will be queued before we have to complete it and
2369 do our thinktime. In other words, this setting effectively caps
2370 the queue depth if the latter is larger.
2371 thinktime_blocks_type=str
2373 Only valid if thinktime is set - control how thinktime_blocks
2374 triggers. The default is `complete', which triggers thinktime
2375 when fio completes thinktime_blocks blocks. If this is set to
2376 `issue', then the trigger happens at the issue side.
2377 thinktime_iotime=time
2379 Only valid if thinktime is set - control thinktime interval by
2380 time. The thinktime stall is repeated after IOs are executed
2381 for thinktime_iotime. For example, `--thinktime_iotime=9s
2382 --thinktime=1s' repeat 10-second cycle with IOs for 9 seconds
2383 and stall for 1 second. When the unit is omitted, thinktime_io‐
2384 time is interpreted as a number of seconds. If this option is
2385 used together with thinktime_blocks, the thinktime stall is re‐
2386 peated after thinktime_iotime or after thinktime_blocks IOs,
2387 whichever happens first.
2388 rate=int[,int][,int]
2391 Cap the bandwidth used by this job. The number is in bytes/sec,
2392 the normal suffix rules apply. Comma-separated values may be
2393 specified for reads, writes, and trims as described in block‐
2394 size.
2395 For example, using `rate=1m,500k' would limit reads to 1MiB/sec
2397 and writes to 500KiB/sec. Capping only reads or writes can be
2398 done with `rate=,500k' or `rate=500k,' where the former will
2399 only limit writes (to 500KiB/sec) and the latter will only limit
2400 reads.
2401 rate_min=int[,int][,int]
2403 Tell fio to do whatever it can to maintain at least this band‐
2404 width. Failing to meet this requirement will cause the job to
2405 exit. Comma-separated values may be specified for reads, writes,
2406 and trims as described in blocksize.
2407 rate_iops=int[,int][,int]
2409 Cap the bandwidth to this number of IOPS. Basically the same as
2410 rate, just specified independently of bandwidth. If the job is
2411 given a block size range instead of a fixed value, the smallest
2412 block size is used as the metric. Comma-separated values may be
2413 specified for reads, writes, and trims as described in block‐
2414 size.
2415 rate_iops_min=int[,int][,int]
2417 If fio doesn't meet this rate of I/O, it will cause the job to
2418 exit. Comma-separated values may be specified for reads,
2419 writes, and trims as described in blocksize.
2420 rate_process=str
2422 This option controls how fio manages rated I/O submissions. The
2423 default is `linear', which submits I/O in a linear fashion with
2424 fixed delays between I/Os that gets adjusted based on I/O com‐
2425 pletion rates. If this is set to `poisson', fio will submit I/O
2426 based on a more real world random request flow, known as the
2427 Poisson process (https://en.wikipedia.org/wiki/Pois‐
2428 son_point_process). The lambda will be 10^6 / IOPS for the given
2429 workload.
2430 rate_ignore_thinktime=bool
2432 By default, fio will attempt to catch up to the specified rate
2433 setting, if any kind of thinktime setting was used. If this op‐
2434 tion is set, then fio will ignore the thinktime and continue do‐
2435 ing IO at the specified rate, instead of entering a catch-up
2436 mode after thinktime is done.
2437 I/O latency
2439 latency_target=time
2440 If set, fio will attempt to find the max performance point that
2441 the given workload will run at while maintaining a latency below
2442 this target. When the unit is omitted, the value is interpreted
2443 in microseconds. See latency_window and latency_percentile.
2444 latency_window=time
2446 Used with latency_target to specify the sample window that the
2447 job is run at varying queue depths to test the performance. When
2448 the unit is omitted, the value is interpreted in microseconds.
2449 latency_percentile=float
2451 The percentage of I/Os that must fall within the criteria speci‐
2452 fied by latency_target and latency_window. If not set, this de‐
2453 faults to 100.0, meaning that all I/Os must be equal or below to
2454 the value set by latency_target.
2455 latency_run=bool
2457 Used with latency_target. If false (default), fio will find the
2458 highest queue depth that meets latency_target and exit. If true,
2459 fio will continue running and try to meet latency_target by ad‐
2460 justing queue depth.
2461 max_latency=time[,time][,time]
2463 If set, fio will exit the job with an ETIMEDOUT error if it ex‐
2464 ceeds this maximum latency. When the unit is omitted, the value
2465 is interpreted in microseconds. Comma-separated values may be
2466 specified for reads, writes, and trims as described in block‐
2467 size.
2468 rate_cycle=int
2470 Average bandwidth for rate and rate_min over this number of mil‐
2471 liseconds. Defaults to 1000.
2472 I/O replay
2474 write_iolog=str
2475 Write the issued I/O patterns to the specified file. See
2476 read_iolog. Specify a separate file for each job, otherwise the
2477 iologs will be interspersed and the file may be corrupt.
2478 read_iolog=str
2480 Open an iolog with the specified filename and replay the I/O
2481 patterns it contains. This can be used to store a workload and
2482 replay it sometime later. The iolog given may also be a blktrace
2483 binary file, which allows fio to replay a workload captured by
2484 blktrace. See blktrace(8) for how to capture such logging data.
2485 For blktrace replay, the file needs to be turned into a blkparse
2486 binary data file first (`blkparse <device> -o /dev/null -d
2487 file_for_fio.bin'). You can specify a number of files by sepa‐
2488 rating the names with a ':' character. See the filename option
2489 for information on how to escape ':' characters within the file
2490 names. These files will be sequentially assigned to job clones
2491 created by numjobs. '-' is a reserved name, meaning read from
2492 stdin, notably if filename is set to '-' which means stdin as
2493 well, then this flag can't be set to '-'.
2494 read_iolog_chunked=bool
2496 Determines how iolog is read. If false (default) entire
2497 read_iolog will be read at once. If selected true, input from
2498 iolog will be read gradually. Useful when iolog is very large,
2499 or it is generated.
2500 merge_blktrace_file=str
2502 When specified, rather than replaying the logs passed to
2503 read_iolog, the logs go through a merge phase which aggregates
2504 them into a single blktrace. The resulting file is then passed
2505 on as the read_iolog parameter. The intention here is to make
2506 the order of events consistent. This limits the influence of the
2507 scheduler compared to replaying multiple blktraces via concur‐
2508 rent jobs.
2509 merge_blktrace_scalars=float_list
2511 This is a percentage based option that is index paired with the
2512 list of files passed to read_iolog. When merging is performed,
2513 scale the time of each event by the corresponding amount. For
2514 example, `--merge_blktrace_scalars="50:100"' runs the first
2515 trace in halftime and the second trace in realtime. This knob is
2516 separately tunable from replay_time_scale which scales the trace
2517 during runtime and will not change the output of the merge un‐
2518 like this option.
2519 merge_blktrace_iters=float_list
2521 This is a whole number option that is index paired with the list
2522 of files passed to read_iolog. When merging is performed, run
2523 each trace for the specified number of iterations. For example,
2524 `--merge_blktrace_iters="2:1"' runs the first trace for two it‐
2525 erations and the second trace for one iteration.
2526 replay_no_stall=bool
2528 When replaying I/O with read_iolog the default behavior is to
2529 attempt to respect the timestamps within the log and replay them
2530 with the appropriate delay between IOPS. By setting this vari‐
2531 able fio will not respect the timestamps and attempt to replay
2532 them as fast as possible while still respecting ordering. The
2533 result is the same I/O pattern to a given device, but different
2534 timings.
2535 replay_time_scale=int
2537 When replaying I/O with read_iolog, fio will honor the original
2538 timing in the trace. With this option, it's possible to scale
2539 the time. It's a percentage option, if set to 50 it means run at
2540 50% the original IO rate in the trace. If set to 200, run at
2541 twice the original IO rate. Defaults to 100.
2542 replay_redirect=str
2544 While replaying I/O patterns using read_iolog the default behav‐
2545 ior is to replay the IOPS onto the major/minor device that each
2546 IOP was recorded from. This is sometimes undesirable because on
2547 a different machine those major/minor numbers can map to a dif‐
2548 ferent device. Changing hardware on the same system can also re‐
2549 sult in a different major/minor mapping. replay_redirect causes
2550 all I/Os to be replayed onto the single specified device regard‐
2551 less of the device it was recorded from. i.e. `replay_redi‐
2552 rect=/dev/sdc' would cause all I/O in the blktrace or iolog to
2553 be replayed onto `/dev/sdc'. This means multiple devices will be
2554 replayed onto a single device, if the trace contains multiple
2555 devices. If you want multiple devices to be replayed concur‐
2556 rently to multiple redirected devices you must blkparse your
2557 trace into separate traces and replay them with independent fio
2558 invocations. Unfortunately this also breaks the strict time or‐
2559 dering between multiple device accesses.
2560 replay_align=int
2562 Force alignment of the byte offsets in a trace to this value.
2563 The value must be a power of 2.
2564 replay_scale=int
2566 Scale bye offsets down by this factor when replaying traces.
2567 Should most likely use replay_align as well.
2568 Threads, processes and job synchronization
2570 replay_skip=str
2571 Sometimes it's useful to skip certain IO types in a replay
2572 trace. This could be, for instance, eliminating the writes in
2573 the trace. Or not replaying the trims/discards, if you are redi‐
2574 recting to a device that doesn't support them. This option
2575 takes a comma separated list of read, write, trim, sync.
2576 thread Fio defaults to creating jobs by using fork, however if this op‐
2578 tion is given, fio will create jobs by using POSIX Threads'
2579 function pthread_create(3) to create threads instead.
2580 wait_for=str
2582 If set, the current job won't be started until all workers of
2583 the specified waitee job are done. wait_for operates on the job
2584 name basis, so there are a few limitations. First, the waitee
2585 must be defined prior to the waiter job (meaning no forward ref‐
2586 erences). Second, if a job is being referenced as a waitee, it
2587 must have a unique name (no duplicate waitees).
2588 nice=int
2590 Run the job with the given nice value. See man nice(2). On Win‐
2591 dows, values less than -15 set the process class to "High"; -1
2592 through -15 set "Above Normal"; 1 through 15 "Below Normal"; and
2593 above 15 "Idle" priority class.
2594 prio=int
2596 Set the I/O priority value of this job. Linux limits us to a
2597 positive value between 0 and 7, with 0 being the highest. See
2598 man ionice(1). Refer to an appropriate manpage for other operat‐
2599 ing systems since meaning of priority may differ. For per-com‐
2600 mand priority setting, see the I/O engine specific `cmdprio_per‐
2601 centage` and `cmdprio` options.
2602 prioclass=int
2604 Set the I/O priority class. See man ionice(1). For per-command
2605 priority setting, see the I/O engine specific `cmdprio_percent‐
2606 age` and `cmdprio_class` options.
2607 cpus_allowed=str
2609 Controls the same options as cpumask, but accepts a textual
2610 specification of the permitted CPUs instead and CPUs are indexed
2611 from 0. So to use CPUs 0 and 5 you would specify `cpus_al‐
2612 lowed=0,5'. This option also allows a range of CPUs to be speci‐
2613 fied -- say you wanted a binding to CPUs 0, 5, and 8 to 15, you
2614 would set `cpus_allowed=0,5,8-15'.
2615 On Windows, when `cpus_allowed' is unset only CPUs from fio's
2617 current processor group will be used and affinity settings are
2618 inherited from the system. An fio build configured to target
2619 Windows 7 makes options that set CPUs processor group aware and
2620 values will set both the processor group and a CPU from within
2621 that group. For example, on a system where processor group 0 has
2622 40 CPUs and processor group 1 has 32 CPUs, `cpus_allowed' values
2623 between 0 and 39 will bind CPUs from processor group 0 and
2624 `cpus_allowed' values between 40 and 71 will bind CPUs from pro‐
2625 cessor group 1. When using `cpus_allowed_policy=shared' all CPUs
2626 specified by a single `cpus_allowed' option must be from the
2627 same processor group. For Windows fio builds not built for Win‐
2628 dows 7, CPUs will only be selected from (and be relative to)
2629 whatever processor group fio happens to be running in and CPUs
2630 from other processor groups cannot be used.
2631 cpus_allowed_policy=str
2633 Set the policy of how fio distributes the CPUs specified by
2634 cpus_allowed or cpumask. Two policies are supported:
2635 shared All jobs will share the CPU set specified.
2637 split Each job will get a unique CPU from the CPU set.
2639 shared is the default behavior, if the option isn't specified.
2641 If split is specified, then fio will assign one cpu per job. If
2642 not enough CPUs are given for the jobs listed, then fio will
2643 roundrobin the CPUs in the set.
2644 cpumask=int
2646 Set the CPU affinity of this job. The parameter given is a bit
2647 mask of allowed CPUs the job may run on. So if you want the al‐
2648 lowed CPUs to be 1 and 5, you would pass the decimal value of (1
2649 << 1 | 1 << 5), or 34. See man sched_setaffinity(2). This may
2650 not work on all supported operating systems or kernel versions.
2651 This option doesn't work well for a higher CPU count than what
2652 you can store in an integer mask, so it can only control cpus
2653 1-32. For boxes with larger CPU counts, use cpus_allowed.
2654 numa_cpu_nodes=str
2656 Set this job running on specified NUMA nodes' CPUs. The argu‐
2657 ments allow comma delimited list of cpu numbers, A-B ranges, or
2658 `all'. Note, to enable NUMA options support, fio must be built
2659 on a system with libnuma-dev(el) installed.
2660 numa_mem_policy=str
2662 Set this job's memory policy and corresponding NUMA nodes. For‐
2663 mat of the arguments:
2664 <mode>[:<nodelist>]
2666 `mode' is one of the following memory policies: `default', `pre‐
2668 fer', `bind', `interleave' or `local'. For `default' and `local'
2669 memory policies, no node needs to be specified. For `prefer',
2670 only one node is allowed. For `bind' and `interleave' the `node‐
2671 list' may be as follows: a comma delimited list of numbers, A-B
2672 ranges, or `all'.
2673 cgroup=str
2675 Add job to this control group. If it doesn't exist, it will be
2676 created. The system must have a mounted cgroup blkio mount point
2677 for this to work. If your system doesn't have it mounted, you
2678 can do so with:
2679 # mount -t cgroup -o blkio none /cgroup
2681 cgroup_weight=int
2683 Set the weight of the cgroup to this value. See the documenta‐
2684 tion that comes with the kernel, allowed values are in the range
2685 of 100..1000.
2686 cgroup_nodelete=bool
2688 Normally fio will delete the cgroups it has created after the
2689 job completion. To override this behavior and to leave cgroups
2690 around after the job completion, set `cgroup_nodelete=1'. This
2691 can be useful if one wants to inspect various cgroup files after
2692 job completion. Default: false.
2693 flow_id=int
2695 The ID of the flow. If not specified, it defaults to being a
2696 global flow. See flow.
2697 flow=int
2699 Weight in token-based flow control. If this value is used, then
2700 fio regulates the activity between two or more jobs sharing the
2701 same flow_id. Fio attempts to keep each job activity propor‐
2702 tional to other jobs' activities in the same flow_id group, with
2703 respect to requested weight per job. That is, if one job has
2704 `flow=3', another job has `flow=2' and another with `flow=1`,
2705 then there will be a roughly 3:2:1 ratio in how much one runs vs
2706 the others.
2707 flow_sleep=int
2709 The period of time, in microseconds, to wait after the flow
2710 counter has exceeded its proportion before retrying operations.
2711 stonewall, wait_for_previous
2713 Wait for preceding jobs in the job file to exit, before starting
2714 this one. Can be used to insert serialization points in the job
2715 file. A stone wall also implies starting a new reporting group,
2716 see group_reporting. Optionally you can use `stonewall=0` to
2717 disable or `stonewall=1` to enable it.
2718 exitall
2720 By default, fio will continue running all other jobs when one
2721 job finishes. Sometimes this is not the desired action. Setting
2722 exitall will instead make fio terminate all jobs in the same
2723 group, as soon as one job of that group finishes.
2724 exit_what=str
2726 By default, fio will continue running all other jobs when one
2727 job finishes. Sometimes this is not the desired action. Setting
2728 exitall will instead make fio terminate all jobs in the same
2729 group. The option exit_what allows you to control which jobs get
2730 terminated when exitall is enabled. The default value is group.
2731 The allowed values are:
2732 all terminates all jobs.
2734 group is the default and does not change the behaviour
2736 of exitall.
2737 stonewall
2739 terminates all currently running jobs across all
2740 groups and continues execution with the next
2741 stonewalled group.
2742 exec_prerun=str
2744 Before running this job, issue the command specified through
2745 system(3). Output is redirected in a file called `jobname.pre‐
2746 run.txt'.
2747 exec_postrun=str
2749 After the job completes, issue the command specified though sys‐
2750 tem(3). Output is redirected in a file called `job‐
2751 name.postrun.txt'.
2752 uid=int
2754 Instead of running as the invoking user, set the user ID to this
2755 value before the thread/process does any work.
2756 gid=int
2758 Set group ID, see uid.
2759 Verification
2761 verify_only
2762 Do not perform specified workload, only verify data still
2763 matches previous invocation of this workload. This option allows
2764 one to check data multiple times at a later date without over‐
2765 writing it. This option makes sense only for workloads that
2766 write data, and does not support workloads with the time_based
2767 option set.
2768 do_verify=bool
2770 Run the verify phase after a write phase. Only valid if verify
2771 is set. Default: true.
2772 verify=str
2774 If writing to a file, fio can verify the file contents after
2775 each iteration of the job. Each verification method also implies
2776 verification of special header, which is written to the begin‐
2777 ning of each block. This header also includes meta information,
2778 like offset of the block, block number, timestamp when block was
2779 written, etc. verify can be combined with verify_pattern option.
2780 The allowed values are:
2781 md5 Use an md5 sum of the data area and store it in
2783 the header of each block.
2784 crc64 Use an experimental crc64 sum of the data area and
2786 store it in the header of each block.
2787 crc32c Use a crc32c sum of the data area and store it in
2789 the header of each block. This will automatically
2790 use hardware acceleration (e.g. SSE4.2 on an x86
2791 or CRC crypto extensions on ARM64) but will fall
2792 back to software crc32c if none is found. Gener‐
2793 ally the fastest checksum fio supports when hard‐
2794 ware accelerated.
2795 crc32c-intel
2797 Synonym for crc32c.
2798 crc32 Use a crc32 sum of the data area and store it in
2800 the header of each block.
2801 crc16 Use a crc16 sum of the data area and store it in
2803 the header of each block.
2804 crc7 Use a crc7 sum of the data area and store it in
2806 the header of each block.
2807 xxhash Use xxhash as the checksum function. Generally the
2809 fastest software checksum that fio supports.
2810 sha512 Use sha512 as the checksum function.
2812 sha256 Use sha256 as the checksum function.
2814 sha1 Use optimized sha1 as the checksum function.
2816 sha3-224
2818 Use optimized sha3-224 as the checksum function.
2819 sha3-256
2821 Use optimized sha3-256 as the checksum function.
2822 sha3-384
2824 Use optimized sha3-384 as the checksum function.
2825 sha3-512
2827 Use optimized sha3-512 as the checksum function.
2828 meta This option is deprecated, since now meta informa‐
2830 tion is included in generic verification header
2831 and meta verification happens by default. For de‐
2832 tailed information see the description of the ver‐
2833 ify setting. This option is kept because of com‐
2834 patibility's sake with old configurations. Do not
2835 use it.
2836 pattern
2838 Verify a strict pattern. Normally fio includes a
2839 header with some basic information and checksum‐
2840 ming, but if this option is set, only the specific
2841 pattern set with verify_pattern is verified.
2842 null Only pretend to verify. Useful for testing inter‐
2844 nals with `ioengine=null', not for much else.
2845 This option can be used for repeated burn-in tests of a system
2847 to make sure that the written data is also correctly read back.
2848 If the data direction given is a read or random read, fio will
2849 assume that it should verify a previously written file. If the
2850 data direction includes any form of write, the verify will be of
2851 the newly written data.
2852 To avoid false verification errors, do not use the norandommap
2854 option when verifying data with async I/O engines and I/O depths
2855 > 1. Or use the norandommap and the lfsr random generator to‐
2856 gether to avoid writing to the same offset with muliple out‐
2857 standing I/Os.
2858 verify_offset=int
2860 Swap the verification header with data somewhere else in the
2861 block before writing. It is swapped back before verifying.
2862 verify_interval=int
2864 Write the verification header at a finer granularity than the
2865 blocksize. It will be written for chunks the size of verify_in‐
2866 terval. blocksize should divide this evenly.
2867 verify_pattern=str
2869 If set, fio will fill the I/O buffers with this pattern. Fio de‐
2870 faults to filling with totally random bytes, but sometimes it's
2871 interesting to fill with a known pattern for I/O verification
2872 purposes. Depending on the width of the pattern, fio will fill
2873 1/2/3/4 bytes of the buffer at the time (it can be either a dec‐
2874 imal or a hex number). The verify_pattern if larger than a
2875 32-bit quantity has to be a hex number that starts with either
2876 "0x" or "0X". Use with verify. Also, verify_pattern supports %o
2877 format, which means that for each block offset will be written
2878 and then verified back, e.g.:
2879 verify_pattern=%o
2881 Or use combination of everything:
2883 verify_pattern=0xff%o"abcd"-12
2885 verify_fatal=bool
2887 Normally fio will keep checking the entire contents before quit‐
2888 ting on a block verification failure. If this option is set, fio
2889 will exit the job on the first observed failure. Default: false.
2890 verify_dump=bool
2892 If set, dump the contents of both the original data block and
2893 the data block we read off disk to files. This allows later
2894 analysis to inspect just what kind of data corruption occurred.
2895 Off by default.
2896 verify_async=int
2898 Fio will normally verify I/O inline from the submitting thread.
2899 This option takes an integer describing how many async offload
2900 threads to create for I/O verification instead, causing fio to
2901 offload the duty of verifying I/O contents to one or more sepa‐
2902 rate threads. If using this offload option, even sync I/O en‐
2903 gines can benefit from using an iodepth setting higher than 1,
2904 as it allows them to have I/O in flight while verifies are run‐
2905 ning. Defaults to 0 async threads, i.e. verification is not
2906 asynchronous.
2907 verify_async_cpus=str
2909 Tell fio to set the given CPU affinity on the async I/O verifi‐
2910 cation threads. See cpus_allowed for the format used.
2911 verify_backlog=int
2913 Fio will normally verify the written contents of a job that uti‐
2914 lizes verify once that job has completed. In other words, every‐
2915 thing is written then everything is read back and verified. You
2916 may want to verify continually instead for a variety of reasons.
2917 Fio stores the meta data associated with an I/O block in memory,
2918 so for large verify workloads, quite a bit of memory would be
2919 used up holding this meta data. If this option is enabled, fio
2920 will write only N blocks before verifying these blocks.
2921 verify_backlog_batch=int
2923 Control how many blocks fio will verify if verify_backlog is
2924 set. If not set, will default to the value of verify_backlog
2925 (meaning the entire queue is read back and verified). If ver‐
2926 ify_backlog_batch is less than verify_backlog then not all
2927 blocks will be verified, if verify_backlog_batch is larger than
2928 verify_backlog, some blocks will be verified more than once.
2929 verify_state_save=bool
2931 When a job exits during the write phase of a verify workload,
2932 save its current state. This allows fio to replay up until that
2933 point, if the verify state is loaded for the verify read phase.
2934 The format of the filename is, roughly:
2935 <type>-<jobname>-<jobindex>-verify.state.
2937 <type> is "local" for a local run, "sock" for a client/server
2939 socket connection, and "ip" (192.168.0.1, for instance) for a
2940 networked client/server connection. Defaults to true.
2941 verify_state_load=bool
2943 If a verify termination trigger was used, fio stores the current
2944 write state of each thread. This can be used at verification
2945 time so that fio knows how far it should verify. Without this
2946 information, fio will run a full verification pass, according to
2947 the settings in the job file used. Default false.
2948 trim_percentage=int
2950 Number of verify blocks to discard/trim.
2951 trim_verify_zero=bool
2953 Verify that trim/discarded blocks are returned as zeros.
2954 trim_backlog=int
2956 Verify that trim/discarded blocks are returned as zeros.
2957 trim_backlog_batch=int
2959 Trim this number of I/O blocks.
2960 experimental_verify=bool
2962 Enable experimental verification.
2963 Steady state
2965 steadystate=str:float, ss=str:float
2966 Define the criterion and limit for assessing steady state per‐
2967 formance. The first parameter designates the criterion whereas
2968 the second parameter sets the threshold. When the criterion
2969 falls below the threshold for the specified duration, the job
2970 will stop. For example, `iops_slope:0.1%' will direct fio to
2971 terminate the job when the least squares regression slope falls
2972 below 0.1% of the mean IOPS. If group_reporting is enabled this
2973 will apply to all jobs in the group. Below is the list of avail‐
2974 able steady state assessment criteria. All assessments are car‐
2975 ried out using only data from the rolling collection window.
2976 Threshold limits can be expressed as a fixed value or as a per‐
2977 centage of the mean in the collection window.
2978 When using this feature, most jobs should include the time_based
2980 and runtime options or the loops option so that fio does not
2981 stop running after it has covered the full size of the specified
2982 file(s) or device(s).
2983 iops Collect IOPS data. Stop the job if all in‐
2985 dividual IOPS measurements are within the
2986 specified limit of the mean IOPS (e.g.,
2987 `iops:2' means that all individual IOPS
2988 values must be within 2 of the mean,
2989 whereas `iops:0.2%' means that all individ‐
2990 ual IOPS values must be within 0.2% of the
2991 mean IOPS to terminate the job).
2992 iops_slope
2994 Collect IOPS data and calculate the least
2995 squares regression slope. Stop the job if
2996 the slope falls below the specified limit.
2997 bw Collect bandwidth data. Stop the job if all
2999 individual bandwidth measurements are
3000 within the specified limit of the mean
3001 bandwidth.
3002 bw_slope
3004 Collect bandwidth data and calculate the
3005 least squares regression slope. Stop the
3006 job if the slope falls below the specified
3007 limit.
3008 steadystate_duration=time, ss_dur=time
3010 A rolling window of this duration will be used to judge
3011 whether steady state has been reached. Data will be col‐
3012 lected once per second. The default is 0 which disables
3013 steady state detection. When the unit is omitted, the
3014 value is interpreted in seconds.
3015 steadystate_ramp_time=time, ss_ramp=time
3017 Allow the job to run for the specified duration before
3018 beginning data collection for checking the steady state
3019 job termination criterion. The default is 0. When the
3020 unit is omitted, the value is interpreted in seconds.
3021 Measurements and reporting
3023 per_job_logs=bool
3024 If set, this generates bw/clat/iops log with per file private
3025 filenames. If not set, jobs with identical names will share the
3026 log filename. Default: true.
3027 group_reporting
3029 It may sometimes be interesting to display statistics for groups
3030 of jobs as a whole instead of for each individual job. This is
3031 especially true if numjobs is used; looking at individual
3032 thread/process output quickly becomes unwieldy. To see the final
3033 report per-group instead of per-job, use group_reporting. Jobs
3034 in a file will be part of the same reporting group, unless if
3035 separated by a stonewall, or by using new_group.
3036 new_group
3038 Start a new reporting group. See: group_reporting. If not given,
3039 all jobs in a file will be part of the same reporting group, un‐
3040 less separated by a stonewall.
3041 stats=bool
3043 By default, fio collects and shows final output results for all
3044 jobs that run. If this option is set to 0, then fio will ignore
3045 it in the final stat output.
3046 write_bw_log=str
3048 If given, write a bandwidth log for this job. Can be used to
3049 store data of the bandwidth of the jobs in their lifetime.
3050 If no str argument is given, the default filename of `job‐
3052 name_type.x.log' is used. Even when the argument is given, fio
3053 will still append the type of log. So if one specifies:
3054 write_bw_log=foo
3056 The actual log name will be `foo_bw.x.log' where `x' is the in‐
3058 dex of the job (1..N, where N is the number of jobs). If
3059 per_job_logs is false, then the filename will not include the
3060 `.x` job index.
3061 The included fio_generate_plots script uses gnuplot to turn
3063 these text files into nice graphs. See the LOG FILE FORMATS sec‐
3064 tion for how data is structured within the file.
3065 write_lat_log=str
3067 Same as write_bw_log, except this option creates I/O submission
3068 (e.g., `name_slat.x.log'), completion (e.g., `name_clat.x.log'),
3069 and total (e.g., `name_lat.x.log') latency files instead. See
3070 write_bw_log for details about the filename format and the LOG
3071 FILE FORMATS section for how data is structured within the
3072 files.
3073 write_hist_log=str
3075 Same as write_bw_log but writes an I/O completion latency his‐
3076 togram file (e.g., `name_hist.x.log') instead. Note that this
3077 file will be empty unless log_hist_msec has also been set. See
3078 write_bw_log for details about the filename format and the LOG
3079 FILE FORMATS section for how data is structured within the file.
3080 write_iops_log=str
3082 Same as write_bw_log, but writes an IOPS file (e.g.
3083 `name_iops.x.log`) instead. Because fio defaults to individual
3084 I/O logging, the value entry in the IOPS log will be 1 unless
3085 windowed logging (see log_avg_msec) has been enabled. See
3086 write_bw_log for details about the filename format and LOG FILE
3087 FORMATS for how data is structured within the file.
3088 log_entries=int
3090 By default, fio will log an entry in the iops, latency, or bw
3091 log for every I/O that completes. The initial number of I/O log
3092 entries is 1024. When the log entries are all used, new log en‐
3093 tries are dynamically allocated. This dynamic log entry alloca‐
3094 tion may negatively impact time-related statistics such as I/O
3095 tail latencies (e.g. 99.9th percentile completion latency). This
3096 option allows specifying a larger initial number of log entries
3097 to avoid run-time allocation of new log entries, resulting in
3098 more precise time-related I/O statistics. Also see log_avg_msec
3099 as well. Defaults to 1024.
3100 log_avg_msec=int
3102 By default, fio will log an entry in the iops, latency, or bw
3103 log for every I/O that completes. When writing to the disk log,
3104 that can quickly grow to a very large size. Setting this option
3105 makes fio average the each log entry over the specified period
3106 of time, reducing the resolution of the log. See log_max_value
3107 as well. Defaults to 0, logging all entries. Also see LOG FILE
3108 FORMATS section.
3109 log_hist_msec=int
3111 Same as log_avg_msec, but logs entries for completion latency
3112 histograms. Computing latency percentiles from averages of in‐
3113 tervals using log_avg_msec is inaccurate. Setting this option
3114 makes fio log histogram entries over the specified period of
3115 time, reducing log sizes for high IOPS devices while retaining
3116 percentile accuracy. See log_hist_coarseness and write_hist_log
3117 as well. Defaults to 0, meaning histogram logging is disabled.
3118 log_hist_coarseness=int
3120 Integer ranging from 0 to 6, defining the coarseness of the res‐
3121 olution of the histogram logs enabled with log_hist_msec. For
3122 each increment in coarseness, fio outputs half as many bins. De‐
3123 faults to 0, for which histogram logs contain 1216 latency bins.
3124 See LOG FILE FORMATS section.
3125 log_max_value=bool
3127 If log_avg_msec is set, fio logs the average over that window.
3128 If you instead want to log the maximum value, set this option to
3129 1. Defaults to 0, meaning that averaged values are logged.
3130 log_offset=bool
3132 If this is set, the iolog options will include the byte offset
3133 for the I/O entry as well as the other data values. Defaults to
3134 0 meaning that offsets are not present in logs. Also see LOG
3135 FILE FORMATS section.
3136 log_prio=bool
3138 If this is set, the iolog options will include the I/O priority
3139 for the I/O entry as well as the other data values. Defaults to
3140 0 meaning that I/O priorities are not present in logs. Also see
3141 LOG FILE FORMATS section.
3142 log_compression=int
3144 If this is set, fio will compress the I/O logs as it goes, to
3145 keep the memory footprint lower. When a log reaches the speci‐
3146 fied size, that chunk is removed and compressed in the back‐
3147 ground. Given that I/O logs are fairly highly compressible, this
3148 yields a nice memory savings for longer runs. The downside is
3149 that the compression will consume some background CPU cycles, so
3150 it may impact the run. This, however, is also true if the log‐
3151 ging ends up consuming most of the system memory. So pick your
3152 poison. The I/O logs are saved normally at the end of a run, by
3153 decompressing the chunks and storing them in the specified log
3154 file. This feature depends on the availability of zlib.
3155 log_compression_cpus=str
3157 Define the set of CPUs that are allowed to handle online log
3158 compression for the I/O jobs. This can provide better isolation
3159 between performance sensitive jobs, and background compression
3160 work. See cpus_allowed for the format used.
3161 log_store_compressed=bool
3163 If set, fio will store the log files in a compressed format.
3164 They can be decompressed with fio, using the --inflate-log com‐
3165 mand line parameter. The files will be stored with a `.fz' suf‐
3166 fix.
3167 log_unix_epoch=bool
3169 If set, fio will log Unix timestamps to the log files produced
3170 by enabling write_type_log for each log type, instead of the de‐
3171 fault zero-based timestamps.
3172 block_error_percentiles=bool
3174 If set, record errors in trim block-sized units from writes and
3175 trims and output a histogram of how many trims it took to get to
3176 errors, and what kind of error was encountered.
3177 bwavgtime=int
3179 Average the calculated bandwidth over the given time. Value is
3180 specified in milliseconds. If the job also does bandwidth log‐
3181 ging through write_bw_log, then the minimum of this option and
3182 log_avg_msec will be used. Default: 500ms.
3183 iopsavgtime=int
3185 Average the calculated IOPS over the given time. Value is speci‐
3186 fied in milliseconds. If the job also does IOPS logging through
3187 write_iops_log, then the minimum of this option and log_avg_msec
3188 will be used. Default: 500ms.
3189 disk_util=bool
3191 Generate disk utilization statistics, if the platform supports
3192 it. Default: true.
3193 disable_lat=bool
3195 Disable measurements of total latency numbers. Useful only for
3196 cutting back the number of calls to gettimeofday(2), as that
3197 does impact performance at really high IOPS rates. Note that to
3198 really get rid of a large amount of these calls, this option
3199 must be used with disable_slat and disable_bw_measurement as
3200 well.
3201 disable_clat=bool
3203 Disable measurements of completion latency numbers. See dis‐
3204 able_lat.
3205 disable_slat=bool
3207 Disable measurements of submission latency numbers. See dis‐
3208 able_lat.
3209 disable_bw_measurement=bool, disable_bw=bool
3211 Disable measurements of throughput/bandwidth numbers. See dis‐
3212 able_lat.
3213 slat_percentiles=bool
3215 Report submission latency percentiles. Submission latency is not
3216 recorded for synchronous ioengines.
3217 clat_percentiles=bool
3219 Report completion latency percentiles.
3220 lat_percentiles=bool
3222 Report total latency percentiles. Total latency is the sum of
3223 submission latency and completion latency.
3224 percentile_list=float_list
3226 Overwrite the default list of percentiles for latencies and the
3227 block error histogram. Each number is a floating point number in
3228 the range (0,100], and the maximum length of the list is 20. Use
3229 ':' to separate the numbers. For example, `--per‐
3230 centile_list=99.5:99.9' will cause fio to report the latency du‐
3231 rations below which 99.5% and 99.9% of the observed latencies
3232 fell, respectively.
3233 significant_figures=int
3235 If using --output-format of `normal', set the significant fig‐
3236 ures to this value. Higher values will yield more precise IOPS
3237 and throughput units, while lower values will round. Requires a
3238 minimum value of 1 and a maximum value of 10. Defaults to 4.
3239 Error handling
3241 exitall_on_error
3242 When one job finishes in error, terminate the rest. The default
3243 is to wait for each job to finish.
3244 continue_on_error=str
3246 Normally fio will exit the job on the first observed failure. If
3247 this option is set, fio will continue the job when there is a
3248 'non-fatal error' (EIO or EILSEQ) until the runtime is exceeded
3249 or the I/O size specified is completed. If this option is used,
3250 there are two more stats that are appended, the total error
3251 count and the first error. The error field given in the stats is
3252 the first error that was hit during the run. The allowed values
3253 are:
3254 none Exit on any I/O or verify errors.
3256 read Continue on read errors, exit on all others.
3258 write Continue on write errors, exit on all others.
3260 io Continue on any I/O error, exit on all others.
3262 verify Continue on verify errors, exit on all others.
3264 all Continue on all errors.
3266 0 Backward-compatible alias for 'none'.
3268 1 Backward-compatible alias for 'all'.
3270 ignore_error=str
3272 Sometimes you want to ignore some errors during test in that
3273 case you can specify error list for each error type, instead of
3274 only being able to ignore the default 'non-fatal error' using
3275 continue_on_error. `ignore_er‐
3276 ror=READ_ERR_LIST,WRITE_ERR_LIST,VERIFY_ERR_LIST' errors for
3277 given error type is separated with ':'. Error may be symbol
3278 ('ENOSPC', 'ENOMEM') or integer. Example:
3279 ignore_error=EAGAIN,ENOSPC:122
3281 This option will ignore EAGAIN from READ, and ENOSPC and
3283 122(EDQUOT) from WRITE. This option works by overriding con‐
3284 tinue_on_error with the list of errors for each error type if
3285 any.
3286 error_dump=bool
3288 If set dump every error even if it is non fatal, true by de‐
3289 fault. If disabled only fatal error will be dumped.
3290 Running predefined workloads
3292 Fio includes predefined profiles that mimic the I/O workloads generated
3293 by other tools.
3294 profile=str
3296 The predefined workload to run. Current profiles are:
3297 tiobench
3299 Threaded I/O bench (tiotest/tiobench) like work‐
3300 load.
3301 act Aerospike Certification Tool (ACT) like workload.
3303 To view a profile's additional options use --cmdhelp after specifying
3305 the profile. For example:
3306 $ fio --profile=act --cmdhelp
3308 Act profile options
3310 device-names=str
3311 Devices to use.
3312 load=int
3314 ACT load multiplier. Default: 1.
3315 test-duration=time
3317 How long the entire test takes to run. When the unit is omitted,
3318 the value is given in seconds. Default: 24h.
3319 threads-per-queue=int
3321 Number of read I/O threads per device. Default: 8.
3322 read-req-num-512-blocks=int
3324 Number of 512B blocks to read at the time. Default: 3.
3325 large-block-op-kbytes=int
3327 Size of large block ops in KiB (writes). Default: 131072.
3328 prep Set to run ACT prep phase.
3330 Tiobench profile options
3332 size=str
3333 Size in MiB.
3334 block=int
3336 Block size in bytes. Default: 4096.
3337 numruns=int
3339 Number of runs.
3340 dir=str
3342 Test directory.
3343 threads=int
3345 Number of threads.
3346
3348 Fio spits out a lot of output. While running, fio will display the sta‐
3349 tus of the jobs created. An example of that would be:
3350 Jobs: 1 (f=1): [_(1),M(1)][24.8%][r=20.5MiB/s,w=23.5MiB/s][r=82,w=94 IOPS][eta 01m:31s]
3352 The characters inside the first set of square brackets denote the cur‐
3354 rent status of each thread. The first character is the first job de‐
3355 fined in the job file, and so forth. The possible values (in typical
3356 life cycle order) are:
3357 P Thread setup, but not started.
3359 C Thread created.
3360 I Thread initialized, waiting or generating necessary data.
3361 p Thread running pre-reading file(s).
3362 / Thread is in ramp period.
3363 R Running, doing sequential reads.
3364 r Running, doing random reads.
3365 W Running, doing sequential writes.
3366 w Running, doing random writes.
3367 M Running, doing mixed sequential reads/writes.
3368 m Running, doing mixed random reads/writes.
3369 D Running, doing sequential trims.
3370 d Running, doing random trims.
3371 F Running, currently waiting for fsync(2).
3372 V Running, doing verification of written data.
3373 f Thread finishing.
3374 E Thread exited, not reaped by main thread yet.
3375 - Thread reaped.
3376 X Thread reaped, exited with an error.
3377 K Thread reaped, exited due to signal.
3378 Fio will condense the thread string as not to take up more space on the
3380 command line than needed. For instance, if you have 10 readers and 10
3381 writers running, the output would look like this:
3382 Jobs: 20 (f=20): [R(10),W(10)][4.0%][r=20.5MiB/s,w=23.5MiB/s][r=82,w=94 IOPS][eta 57m:36s]
3384 Note that the status string is displayed in order, so it's possible to
3386 tell which of the jobs are currently doing what. In the example above
3387 this means that jobs 1--10 are readers and 11--20 are writers.
3388 The other values are fairly self explanatory -- number of threads cur‐
3390 rently running and doing I/O, the number of currently open files (f=),
3391 the estimated completion percentage, the rate of I/O since last check
3392 (read speed listed first, then write speed and optionally trim speed)
3393 in terms of bandwidth and IOPS, and time to completion for the current
3394 running group. It's impossible to estimate runtime of the following
3395 groups (if any).
3396 When fio is done (or interrupted by Ctrl-C), it will show the data for
3398 each thread, group of threads, and disks in that order. For each over‐
3399 all thread (or group) the output looks like:
3400 Client1: (groupid=0, jobs=1): err= 0: pid=16109: Sat Jun 24 12:07:54 2017
3402 write: IOPS=88, BW=623KiB/s (638kB/s)(30.4MiB/50032msec)
3403 slat (nsec): min=500, max=145500, avg=8318.00, stdev=4781.50
3404 clat (usec): min=170, max=78367, avg=4019.02, stdev=8293.31
3405 lat (usec): min=174, max=78375, avg=4027.34, stdev=8291.79
3406 clat percentiles (usec):
3407 | 1.00th=[ 302], 5.00th=[ 326], 10.00th=[ 343], 20.00th=[ 363],
3408 | 30.00th=[ 392], 40.00th=[ 404], 50.00th=[ 416], 60.00th=[ 445],
3409 | 70.00th=[ 816], 80.00th=[ 6718], 90.00th=[12911], 95.00th=[21627],
3410 | 99.00th=[43779], 99.50th=[51643], 99.90th=[68682], 99.95th=[72877],
3411 | 99.99th=[78119]
3412 bw ( KiB/s): min= 532, max= 686, per=0.10%, avg=622.87, stdev=24.82, samples= 100
3413 iops : min= 76, max= 98, avg=88.98, stdev= 3.54, samples= 100
3414 lat (usec) : 250=0.04%, 500=64.11%, 750=4.81%, 1000=2.79%
3415 lat (msec) : 2=4.16%, 4=1.84%, 10=4.90%, 20=11.33%, 50=5.37%
3416 lat (msec) : 100=0.65%
3417 cpu : usr=0.27%, sys=0.18%, ctx=12072, majf=0, minf=21
3418 IO depths : 1=85.0%, 2=13.1%, 4=1.8%, 8=0.1%, 16=0.0%, 32=0.0%, >=64=0.0%
3419 submit : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
3420 complete : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
3421 issued rwt: total=0,4450,0, short=0,0,0, dropped=0,0,0
3422 latency : target=0, window=0, percentile=100.00%, depth=8
3423 The job name (or first job's name when using group_reporting) is
3425 printed, along with the group id, count of jobs being aggregated, last
3426 error id seen (which is 0 when there are no errors), pid/tid of that
3427 thread and the time the job/group completed. Below are the I/O statis‐
3428 tics for each data direction performed (showing writes in the example
3429 above). In the order listed, they denote:
3430 read/write/trim
3432 The string before the colon shows the I/O direction the
3433 statistics are for. IOPS is the average I/Os performed
3434 per second. BW is the average bandwidth rate shown as:
3435 value in power of 2 format (value in power of 10 format).
3436 The last two values show: (total I/O performed in power
3437 of 2 format / runtime of that thread).
3438 slat Submission latency (min being the minimum, max being the
3440 maximum, avg being the average, stdev being the standard
3441 deviation). This is the time it took to submit the I/O.
3442 For sync I/O this row is not displayed as the slat is re‐
3443 ally the completion latency (since queue/complete is one
3444 operation there). This value can be in nanoseconds, mi‐
3445 croseconds or milliseconds --- fio will choose the most
3446 appropriate base and print that (in the example above
3447 nanoseconds was the best scale). Note: in --minimal mode
3448 latencies are always expressed in microseconds.
3449 clat Completion latency. Same names as slat, this denotes the
3451 time from submission to completion of the I/O pieces. For
3452 sync I/O, clat will usually be equal (or very close) to
3453 0, as the time from submit to complete is basically just
3454 CPU time (I/O has already been done, see slat explana‐
3455 tion).
3456 lat Total latency. Same names as slat and clat, this denotes
3458 the time from when fio created the I/O unit to completion
3459 of the I/O operation.
3460 bw Bandwidth statistics based on samples. Same names as the
3462 xlat stats, but also includes the number of samples taken
3463 (samples) and an approximate percentage of total aggre‐
3464 gate bandwidth this thread received in its group (per).
3465 This last value is only really useful if the threads in
3466 this group are on the same disk, since they are then com‐
3467 peting for disk access.
3468 iops IOPS statistics based on samples. Same names as bw.
3470 lat (nsec/usec/msec)
3472 The distribution of I/O completion latencies. This is the
3473 time from when I/O leaves fio and when it gets completed.
3474 Unlike the separate read/write/trim sections above, the
3475 data here and in the remaining sections apply to all I/Os
3476 for the reporting group. 250=0.04% means that 0.04% of
3477 the I/Os completed in under 250us. 500=64.11% means that
3478 64.11% of the I/Os required 250 to 499us for completion.
3479 cpu CPU usage. User and system time, along with the number of
3481 context switches this thread went through, usage of sys‐
3482 tem and user time, and finally the number of major and
3483 minor page faults. The CPU utilization numbers are aver‐
3484 ages for the jobs in that reporting group, while the con‐
3485 text and fault counters are summed.
3486 IO depths
3488 The distribution of I/O depths over the job lifetime. The
3489 numbers are divided into powers of 2 and each entry cov‐
3490 ers depths from that value up to those that are lower
3491 than the next entry -- e.g., 16= covers depths from 16 to
3492 31. Note that the range covered by a depth distribution
3493 entry can be different to the range covered by the equiv‐
3494 alent submit/complete distribution entry.
3495 IO submit
3497 How many pieces of I/O were submitting in a single submit
3498 call. Each entry denotes that amount and below, until the
3499 previous entry -- e.g., 16=100% means that we submitted
3500 anywhere between 9 to 16 I/Os per submit call. Note that
3501 the range covered by a submit distribution entry can be
3502 different to the range covered by the equivalent depth
3503 distribution entry.
3504 IO complete
3506 Like the above submit number, but for completions in‐
3507 stead.
3508 IO issued rwt
3510 The number of read/write/trim requests issued, and how
3511 many of them were short or dropped.
3512 IO latency
3514 These values are for latency_target and related options.
3515 When these options are engaged, this section describes
3516 the I/O depth required to meet the specified latency tar‐
3517 get.
3518 After each client has been listed, the group statistics are printed.
3520 They will look like this:
3521 Run status group 0 (all jobs):
3523 READ: bw=20.9MiB/s (21.9MB/s), 10.4MiB/s-10.8MiB/s (10.9MB/s-11.3MB/s), io=64.0MiB (67.1MB), run=2973-3069msec
3524 WRITE: bw=1231KiB/s (1261kB/s), 616KiB/s-621KiB/s (630kB/s-636kB/s), io=64.0MiB (67.1MB), run=52747-53223msec
3525 For each data direction it prints:
3527 bw Aggregate bandwidth of threads in this group followed by
3529 the minimum and maximum bandwidth of all the threads in
3530 this group. Values outside of brackets are power-of-2
3531 format and those within are the equivalent value in a
3532 power-of-10 format.
3533 io Aggregate I/O performed of all threads in this group. The
3535 format is the same as bw.
3536 run The smallest and longest runtimes of the threads in this
3538 group.
3539 And finally, the disk statistics are printed. This is Linux specific.
3541 They will look like this:
3542 Disk stats (read/write):
3544 sda: ios=16398/16511, merge=30/162, ticks=6853/819634, in_queue=826487, util=100.00%
3545 Each value is printed for both reads and writes, with reads first. The
3547 numbers denote:
3548 ios Number of I/Os performed by all groups.
3550 merge Number of merges performed by the I/O scheduler.
3552 ticks Number of ticks we kept the disk busy.
3554 in_queue
3556 Total time spent in the disk queue.
3557 util The disk utilization. A value of 100% means we kept the
3559 disk busy constantly, 50% would be a disk idling half of
3560 the time.
3561 It is also possible to get fio to dump the current output while it is
3563 running, without terminating the job. To do that, send fio the USR1
3564 signal. You can also get regularly timed dumps by using the --sta‐
3565 tus-interval parameter, or by creating a file in `/tmp' named
3566 `fio-dump-status'. If fio sees this file, it will unlink it and dump
3567 the current output status.
3568
3570 For scripted usage where you typically want to generate tables or
3571 graphs of the results, fio can output the results in a semicolon sepa‐
3572 rated format. The format is one long line of values, such as:
3573 2;card0;0;0;7139336;121836;60004;1;10109;27.932460;116.933948;220;126861;3495.446807;1085.368601;226;126864;3523.635629;1089.012448;24063;99944;50.275485%;59818.274627;5540.657370;7155060;122104;60004;1;8338;29.086342;117.839068;388;128077;5032.488518;1234.785715;391;128085;5061.839412;1236.909129;23436;100928;50.287926%;59964.832030;5644.844189;14.595833%;19.394167%;123706;0;7313;0.1%;0.1%;0.1%;0.1%;0.1%;0.1%;100.0%;0.00%;0.00%;0.00%;0.00%;0.00%;0.00%;0.01%;0.02%;0.05%;0.16%;6.04%;40.40%;52.68%;0.64%;0.01%;0.00%;0.01%;0.00%;0.00%;0.00%;0.00%;0.00%
3575 A description of this job goes here.
3576 The job description (if provided) follows on a second line for terse
3578 v2. It appears on the same line for other terse versions.
3579 To enable terse output, use the --minimal or `--output-format=terse'
3581 command line options. The first value is the version of the terse out‐
3582 put format. If the output has to be changed for some reason, this num‐
3583 ber will be incremented by 1 to signify that change.
3584 Split up, the format is as follows (comments in brackets denote when a
3586 field was introduced or whether it's specific to some terse version):
3587 terse version, fio version [v3], jobname, groupid, error
3589 READ status:
3591 Total IO (KiB), bandwidth (KiB/sec), IOPS, runtime (msec)
3593 Submission latency: min, max, mean, stdev (usec)
3594 Completion latency: min, max, mean, stdev (usec)
3595 Completion latency percentiles: 20 fields (see below)
3596 Total latency: min, max, mean, stdev (usec)
3597 Bw (KiB/s): min, max, aggregate percentage of total, mean, stdev, number of samples [v5]
3598 IOPS [v5]: min, max, mean, stdev, number of samples
3599 WRITE status:
3601 Total IO (KiB), bandwidth (KiB/sec), IOPS, runtime (msec)
3603 Submission latency: min, max, mean, stdev (usec)
3604 Completion latency: min, max, mean, stdev (usec)
3605 Completion latency percentiles: 20 fields (see below)
3606 Total latency: min, max, mean, stdev (usec)
3607 Bw (KiB/s): min, max, aggregate percentage of total, mean, stdev, number of samples [v5]
3608 IOPS [v5]: min, max, mean, stdev, number of samples
3609 TRIM status [all but version 3]:
3611 Fields are similar to READ/WRITE status.
3613 CPU usage:
3615 user, system, context switches, major faults, minor faults
3617 I/O depths:
3619 <=1, 2, 4, 8, 16, 32, >=64
3621 I/O latencies microseconds:
3623 <=2, 4, 10, 20, 50, 100, 250, 500, 750, 1000
3625 I/O latencies milliseconds:
3627 <=2, 4, 10, 20, 50, 100, 250, 500, 750, 1000, 2000, >=2000
3629 Disk utilization [v3]:
3631 disk name, read ios, write ios, read merges, write merges, read ticks, write ticks, time spent in queue, disk utilization percentage
3633 Additional Info (dependent on continue_on_error, default off):
3635 total # errors, first error code
3637 Additional Info (dependent on description being set):
3639 Text description
3641 Completion latency percentiles can be a grouping of up to 20 sets, so
3643 for the terse output fio writes all of them. Each field will look like
3644 this:
3645 1.00%=6112
3647 which is the Xth percentile, and the `usec' latency associated with it.
3649 For Disk utilization, all disks used by fio are shown. So for each disk
3651 there will be a disk utilization section.
3652 Below is a single line containing short names for each of the fields in
3654 the minimal output v3, separated by semicolons:
3655 terse_version_3;fio_version;jobname;groupid;error;read_kb;read_bandwidth_kb;read_iops;read_runtime_ms;read_slat_min_us;read_slat_max_us;read_slat_mean_us;read_slat_dev_us;read_clat_min_us;read_clat_max_us;read_clat_mean_us;read_clat_dev_us;read_clat_pct01;read_clat_pct02;read_clat_pct03;read_clat_pct04;read_clat_pct05;read_clat_pct06;read_clat_pct07;read_clat_pct08;read_clat_pct09;read_clat_pct10;read_clat_pct11;read_clat_pct12;read_clat_pct13;read_clat_pct14;read_clat_pct15;read_clat_pct16;read_clat_pct17;read_clat_pct18;read_clat_pct19;read_clat_pct20;read_tlat_min_us;read_lat_max_us;read_lat_mean_us;read_lat_dev_us;read_bw_min_kb;read_bw_max_kb;read_bw_agg_pct;read_bw_mean_kb;read_bw_dev_kb;write_kb;write_bandwidth_kb;write_iops;write_runtime_ms;write_slat_min_us;write_slat_max_us;write_slat_mean_us;write_slat_dev_us;write_clat_min_us;write_clat_max_us;write_clat_mean_us;write_clat_dev_us;write_clat_pct01;write_clat_pct02;write_clat_pct03;write_clat_pct04;write_clat_pct05;write_clat_pct06;write_clat_pct07;write_clat_pct08;write_clat_pct09;write_clat_pct10;write_clat_pct11;write_clat_pct12;write_clat_pct13;write_clat_pct14;write_clat_pct15;write_clat_pct16;write_clat_pct17;write_clat_pct18;write_clat_pct19;write_clat_pct20;write_tlat_min_us;write_lat_max_us;write_lat_mean_us;write_lat_dev_us;write_bw_min_kb;write_bw_max_kb;write_bw_agg_pct;write_bw_mean_kb;write_bw_dev_kb;cpu_user;cpu_sys;cpu_csw;cpu_mjf;cpu_minf;iodepth_1;iodepth_2;iodepth_4;iodepth_8;iodepth_16;iodepth_32;iodepth_64;lat_2us;lat_4us;lat_10us;lat_20us;lat_50us;lat_100us;lat_250us;lat_500us;lat_750us;lat_1000us;lat_2ms;lat_4ms;lat_10ms;lat_20ms;lat_50ms;lat_100ms;lat_250ms;lat_500ms;lat_750ms;lat_1000ms;lat_2000ms;lat_over_2000ms;disk_name;disk_read_iops;disk_write_iops;disk_read_merges;disk_write_merges;disk_read_ticks;write_ticks;disk_queue_time;disk_util
3657 In client/server mode terse output differs from what appears when jobs
3659 are run locally. Disk utilization data is omitted from the standard
3660 terse output and for v3 and later appears on its own separate line at
3661 the end of each terse reporting cycle.
3662
3664 The json output format is intended to be both human readable and conve‐
3665 nient for automated parsing. For the most part its sections mirror
3666 those of the normal output. The runtime value is reported in msec and
3667 the bw value is reported in 1024 bytes per second units.
3668
3670 The json+ output format is identical to the json output format except
3671 that it adds a full dump of the completion latency bins. Each bins ob‐
3672 ject contains a set of (key, value) pairs where keys are latency dura‐
3673 tions and values count how many I/Os had completion latencies of the
3674 corresponding duration. For example, consider:
3675 "bins" : { "87552" : 1, "89600" : 1, "94720" : 1, "96768" : 1,
3677 "97792" : 1, "99840" : 1, "100864" : 2, "103936" : 6, "104960" :
3678 534, "105984" : 5995, "107008" : 7529, ... }
3679 This data indicates that one I/O required 87,552ns to complete, two
3681 I/Os required 100,864ns to complete, and 7529 I/Os required 107,008ns
3682 to complete.
3683 Also included with fio is a Python script fio_jsonplus_clat2csv that
3685 takes json+ output and generates CSV-formatted latency data suitable
3686 for plotting.
3687 The latency durations actually represent the midpoints of latency in‐
3689 tervals. For details refer to `stat.h' in the fio source.
3690
3692 There are two trace file format that you can encounter. The older (v1)
3693 format is unsupported since version 1.20-rc3 (March 2008). It will
3694 still be described below in case that you get an old trace and want to
3695 understand it.
3696 In any case the trace is a simple text file with a single action per
3698 line.
3699 Trace file format v1
3701 Each line represents a single I/O action in the following for‐
3702 mat:
3703 rw, offset, length
3705 where `rw=0/1' for read/write, and the `offset' and `length' en‐
3707 tries being in bytes.
3708 This format is not supported in fio versions >= 1.20-rc3.
3710 Trace file format v2
3712 The second version of the trace file format was added in fio
3713 version 1.17. It allows to access more then one file per trace
3714 and has a bigger set of possible file actions.
3715 The first line of the trace file has to be:
3717 "fio version 2 iolog"
3719 Following this can be lines in two different formats, which are
3721 described below.
3722 The file management format:
3724 filename action
3725 The `filename' is given as an absolute path. The `action'
3727 can be one of these:
3728 add Add the given `filename' to the trace.
3730 open Open the file with the given `filename'.
3732 The `filename' has to have been added with
3733 the add action before.
3734 close Close the file with the given `filename'.
3736 The file has to have been opened before.
3737 The file I/O action format:
3739 filename action offset length
3740 The `filename' is given as an absolute path, and has to
3742 have been added and opened before it can be used with
3743 this format. The `offset' and `length' are given in
3744 bytes. The `action' can be one of these:
3745 wait Wait for `offset' microseconds. Everything
3747 below 100 is discarded. The time is rela‐
3748 tive to the previous `wait' statement.
3749 read Read `length' bytes beginning from `off‐
3751 set'.
3752 write Write `length' bytes beginning from `off‐
3754 set'.
3755 sync fsync(2) the file.
3757 datasync
3759 fdatasync(2) the file.
3760 trim Trim the given file from the given `offset'
3762 for `length' bytes.
3763
3765 Colocation is a common practice used to get the most out of a machine.
3766 Knowing which workloads play nicely with each other and which ones
3767 don't is a much harder task. While fio can replay workloads concur‐
3768 rently via multiple jobs, it leaves some variability up to the sched‐
3769 uler making results harder to reproduce. Merging is a way to make the
3770 order of events consistent.
3771 Merging is integrated into I/O replay and done when a merge_blk‐
3773 trace_file is specified. The list of files passed to read_iolog go
3774 through the merge process and output a single file stored to the speci‐
3775 fied file. The output file is passed on as if it were the only file
3776 passed to read_iolog. An example would look like:
3777 $ fio --read_iolog="<file1>:<file2>" --merge_blk‐
3779 trace_file="<output_file>"
3780 Creating only the merged file can be done by passing the command line
3782 argument merge-blktrace-only.
3783 Scaling traces can be done to see the relative impact of any particular
3785 trace being slowed down or sped up. merge_blktrace_scalars takes in a
3786 colon separated list of percentage scalars. It is index paired with the
3787 files passed to read_iolog.
3788 With scaling, it may be desirable to match the running time of all
3790 traces. This can be done with merge_blktrace_iters. It is index paired
3791 with read_iolog just like merge_blktrace_scalars.
3792 In an example, given two traces, A and B, each 60s long. If we want to
3794 see the impact of trace A issuing IOs twice as fast and repeat trace A
3795 over the runtime of trace B, the following can be done:
3796 $ fio --read_iolog="<trace_a>:"<trace_b>" --merge_blk‐
3798 trace_file"<output_file>" --merge_blktrace_scalars="50:100"
3799 --merge_blktrace_iters="2:1"
3800 This runs trace A at 2x the speed twice for approximately the same run‐
3802 time as a single run of trace B.
3803
3805 In some cases, we want to understand CPU overhead in a test. For exam‐
3806 ple, we test patches for the specific goodness of whether they reduce
3807 CPU usage. Fio implements a balloon approach to create a thread per
3808 CPU that runs at idle priority, meaning that it only runs when nobody
3809 else needs the cpu. By measuring the amount of work completed by the
3810 thread, idleness of each CPU can be derived accordingly.
3811 An unit work is defined as touching a full page of unsigned characters.
3813 Mean and standard deviation of time to complete an unit work is re‐
3814 ported in "unit work" section. Options can be chosen to report detailed
3815 percpu idleness or overall system idleness by aggregating percpu stats.
3816
3818 Fio is usually run in one of two ways, when data verification is done.
3819 The first is a normal write job of some sort with verify enabled. When
3820 the write phase has completed, fio switches to reads and verifies ev‐
3821 erything it wrote. The second model is running just the write phase,
3822 and then later on running the same job (but with reads instead of
3823 writes) to repeat the same I/O patterns and verify the contents. Both
3824 of these methods depend on the write phase being completed, as fio oth‐
3825 erwise has no idea how much data was written.
3826 With verification triggers, fio supports dumping the current write
3828 state to local files. Then a subsequent read verify workload can load
3829 this state and know exactly where to stop. This is useful for testing
3830 cases where power is cut to a server in a managed fashion, for in‐
3831 stance.
3832 A verification trigger consists of two things:
3834 1) Storing the write state of each job.
3836 2) Executing a trigger command.
3838 The write state is relatively small, on the order of hundreds of bytes
3840 to single kilobytes. It contains information on the number of comple‐
3841 tions done, the last X completions, etc.
3842 A trigger is invoked either through creation ('touch') of a specified
3844 file in the system, or through a timeout setting. If fio is run with
3845 `--trigger-file=/tmp/trigger-file', then it will continually check for
3846 the existence of `/tmp/trigger-file'. When it sees this file, it will
3847 fire off the trigger (thus saving state, and executing the trigger com‐
3848 mand).
3849 For client/server runs, there's both a local and remote trigger. If fio
3851 is running as a server backend, it will send the job states back to the
3852 client for safe storage, then execute the remote trigger, if specified.
3853 If a local trigger is specified, the server will still send back the
3854 write state, but the client will then execute the trigger.
3855 Verification trigger example
3857 Let's say we want to run a powercut test on the remote Linux ma‐
3858 chine 'server'. Our write workload is in `write-test.fio'. We
3859 want to cut power to 'server' at some point during the run, and
3860 we'll run this test from the safety or our local machine, 'lo‐
3861 calbox'. On the server, we'll start the fio backend normally:
3862 server# fio --server
3864 and on the client, we'll fire off the workload:
3866 localbox$ fio --client=server --trig‐
3868 ger-file=/tmp/my-trigger --trigger-remote="bash -c "echo
3869 b > /proc/sysrq-triger""
3870 We set `/tmp/my-trigger' as the trigger file, and we tell fio to
3872 execute:
3873 echo b > /proc/sysrq-trigger
3875 on the server once it has received the trigger and sent us the
3877 write state. This will work, but it's not really cutting power
3878 to the server, it's merely abruptly rebooting it. If we have a
3879 remote way of cutting power to the server through IPMI or simi‐
3880 lar, we could do that through a local trigger command instead.
3881 Let's assume we have a script that does IPMI reboot of a given
3882 hostname, ipmi-reboot. On localbox, we could then have run fio
3883 with a local trigger instead:
3884 localbox$ fio --client=server --trig‐
3886 ger-file=/tmp/my-trigger --trigger="ipmi-reboot server"
3887 For this case, fio would wait for the server to send us the
3889 write state, then execute `ipmi-reboot server' when that hap‐
3890 pened.
3891 Loading verify state
3893 To load stored write state, a read verification job file must
3894 contain the verify_state_load option. If that is set, fio will
3895 load the previously stored state. For a local fio run this is
3896 done by loading the files directly, and on a client/server run,
3897 the server backend will ask the client to send the files over
3898 and load them from there.
3899
3901 Fio supports a variety of log file formats, for logging latencies,
3902 bandwidth, and IOPS. The logs share a common format, which looks like
3903 this:
3904 time (msec), value, data direction, block size (bytes), offset
3906 (bytes), command priority
3907 `Time' for the log entry is always in milliseconds. The `value' logged
3909 depends on the type of log, it will be one of the following:
3910 Latency log
3912 Value is latency in nsecs
3913 Bandwidth log
3915 Value is in KiB/sec
3916 IOPS log
3918 Value is IOPS
3919 `Data direction' is one of the following:
3921 0 I/O is a READ
3923 1 I/O is a WRITE
3925 2 I/O is a TRIM
3927 The entry's `block size' is always in bytes. The `offset' is the posi‐
3929 tion in bytes from the start of the file for that particular I/O. The
3930 logging of the offset can be toggled with log_offset.
3931 If log_prio is not set, the entry's `Command priority` is 1 for an IO
3933 executed with the highest RT priority class (prioclass=1 or cmd‐
3934 prio_class=1) and 0 otherwise. This is controlled by the prioclass op‐
3935 tion and the ioengine specific cmdprio_percentage cmdprio_class op‐
3936 tions. If log_prio is set, the entry's `Command priority` is the prior‐
3937 ity set for the IO, as a 16-bits hexadecimal number with the lowest 13
3938 bits indicating the priority value (prio and cmdprio options) and the
3939 highest 3 bits indicating the IO priority class (prioclass and cmd‐
3940 prio_class options).
3941 Fio defaults to logging every individual I/O but when windowed logging
3943 is set through log_avg_msec, either the average (by default) or the
3944 maximum (log_max_value is set) `value' seen over the specified period
3945 of time is recorded. Each `data direction' seen within the window pe‐
3946 riod will aggregate its values in a separate row. Further, when using
3947 windowed logging the `block size' and `offset' entries will always con‐
3948 tain 0.
3949
3951 Normally fio is invoked as a stand-alone application on the machine
3952 where the I/O workload should be generated. However, the backend and
3953 frontend of fio can be run separately i.e., the fio server can generate
3954 an I/O workload on the "Device Under Test" while being controlled by a
3955 client on another machine.
3956 Start the server on the machine which has access to the storage DUT:
3958 $ fio --server=args
3960 where `args' defines what fio listens to. The arguments are of the form
3962 `type,hostname' or `IP,port'. `type' is either `ip' (or ip4) for TCP/IP
3963 v4, `ip6' for TCP/IP v6, or `sock' for a local unix domain socket.
3964 `hostname' is either a hostname or IP address, and `port' is the port
3965 to listen to (only valid for TCP/IP, not a local socket). Some exam‐
3966 ples:
3967 1) fio --server
3969 Start a fio server, listening on all interfaces on the
3970 default port (8765).
3971 2) fio --server=ip:hostname,4444
3973 Start a fio server, listening on IP belonging to hostname
3974 and on port 4444.
3975 3) fio --server=ip6:::1,4444
3977 Start a fio server, listening on IPv6 localhost ::1 and
3978 on port 4444.
3979 4) fio --server=,4444
3981 Start a fio server, listening on all interfaces on port
3982 4444.
3983 5) fio --server=1.2.3.4
3985 Start a fio server, listening on IP 1.2.3.4 on the de‐
3986 fault port.
3987 6) fio --server=sock:/tmp/fio.sock
3989 Start a fio server, listening on the local socket
3990 `/tmp/fio.sock'.
3991 Once a server is running, a "client" can connect to the fio server
3993 with:
3994 $ fio <local-args> --client=<server> <remote-args> <job file(s)>
3996 where `local-args' are arguments for the client where it is running,
3998 `server' is the connect string, and `remote-args' and `job file(s)' are
3999 sent to the server. The `server' string follows the same format as it
4000 does on the server side, to allow IP/hostname/socket and port strings.
4001 Fio can connect to multiple servers this way:
4003 $ fio --client=<server1> <job file(s)> --client=<server2> <job
4005 file(s)>
4006 If the job file is located on the fio server, then you can tell the
4008 server to load a local file as well. This is done by using --re‐
4009 mote-config:
4010 $ fio --client=server --remote-config /path/to/file.fio
4012 Then fio will open this local (to the server) job file instead of being
4014 passed one from the client.
4015 If you have many servers (example: 100 VMs/containers), you can input a
4017 pathname of a file containing host IPs/names as the parameter value for
4018 the --client option. For example, here is an example `host.list' file
4019 containing 2 hostnames:
4020 host1.your.dns.domain
4022 host2.your.dns.domain
4023 The fio command would then be:
4025 $ fio --client=host.list <job file(s)>
4027 In this mode, you cannot input server-specific parameters or job files
4029 -- all servers receive the same job file.
4030 In order to let `fio --client' runs use a shared filesystem from multi‐
4032 ple hosts, `fio --client' now prepends the IP address of the server to
4033 the filename. For example, if fio is using the directory `/mnt/nfs/fio'
4034 and is writing filename `fileio.tmp', with a --client `hostfile' con‐
4035 taining two hostnames `h1' and `h2' with IP addresses 192.168.10.120
4036 and 192.168.10.121, then fio will create two files:
4037 /mnt/nfs/fio/192.168.10.120.fileio.tmp
4039 /mnt/nfs/fio/192.168.10.121.fileio.tmp
4040 Terse output in client/server mode will differ slightly from what is
4042 produced when fio is run in stand-alone mode. See the terse output sec‐
4043 tion for details.
4044
4046 fio was written by Jens Axboe <axboe@kernel.dk>.
4047 This man page was written by Aaron Carroll <aaronc@cse.unsw.edu.au>
4048 based on documentation by Jens Axboe.
4049 This man page was rewritten by Tomohiro Kusumi <tkusumi@tuxera.com>
4050 based on documentation by Jens Axboe.
4051
4053 Report bugs to the fio mailing list <fio@vger.kernel.org>.
4054 See REPORTING-BUGS.
4055 REPORTING-BUGS: http://git.kernel.dk/cgit/fio/plain/REPORTING-BUGS
4057
4059 For further documentation see HOWTO and README.
4060 Sample jobfiles are available in the `examples/' directory.
4061 These are typically located under `/usr/share/doc/fio'.
4062 HOWTO: http://git.kernel.dk/cgit/fio/plain/HOWTO
4064 README: http://git.kernel.dk/cgit/fio/plain/README
4065User Manual August 2017 fio(1)