1prun(1) PRRTE prun(1)
2
6 prun - Execute serial and parallel jobs with the PMIx Reference Runtime
7 (PRTE).
8
10 prun requires a running prte Distributed Virtual Machine (DVM) to be
11 running at the time of the call. See prte(1) for more information.
12 Single Process Multiple Data (SPMD) Model:
14 prun [ options ] <program> [ <args> ]
16 Multiple Instruction Multiple Data (MIMD) Model:
18 prun [ global_options ] \
20 [ local_options1 ] <program1> [ <args1> ] : \
21 [ local_options2 ] <program2> [ <args2> ] : \
22 ... : \
23 [ local_optionsN ] <programN> [ <argsN> ]
24 Note that in both models, invoking prun via an absolute path name is
26 equivalent to specifying the --prefix option with a <dir> value equiva‐
27 lent to the directory where prun resides, minus its last subdirectory.
28 For example:
29 $ /usr/local/bin/prun ...
31 is equivalent to
33 $ prun --prefix /usr/local
35
37 If you are simply looking for how to run an application, you probably
38 want to use a command line of the following form:
39 $ prun [ -np X ] [ --hostfile <filename> ] <program>
41 This will run X copies of <program> in your current run-time environ‐
43 ment over the set of hosts specified by <filename>, scheduling (by de‐
44 fault) in a round-robin fashion by CPU slot. If running under a sup‐
45 ported resource manager a hostfile is usually not required unless the
46 caller wishes to further restrict the set of resources used for that
47 job.
48 Please note that PRTE automatically binds processes. See prte-map(1)
50 for defaults for the mapping, ranking, and binding of processes.
51 If your application uses threads, then you probably want to ensure that
53 you are either not bound at all (by specifying --bind-to none), or
54 bound to multiple cores using an appropriate binding level or specific
55 number of processing elements per application process.
56 Default ranking is by slot if number of processes <= 2, otherwise de‐
58 fault to ranking by package (formally known as “socket”).
59 See prte-map(1) for more details on mapping, ranking, and binding op‐
61 tions.
62
64 This section includes many commonly used options. There may be other
65 options listed with prun --help.
66 prun will send the name of the directory where it was invoked on the
68 local node to each of the remote nodes, and attempt to change to that
69 directory. See the “Current Working Directory” section below for fur‐
70 ther details.
71 <program>
73 The program executable. This is identified as the first non-
74 recognized argument to prun.
75 <args> Pass these run-time arguments to every new process. These must
77 always be the last arguments to prun after the <program>. If an
78 app context file is used, <args> will be ignored.
79 -h, --help
81 Display help for this command
82 -q, --quiet
84 Suppress informative messages from prun during application exe‐
85 cution.
86 -v, --verbose
88 Be verbose
89 -V, --version
91 Print version number. If no other arguments are given, this
92 will also cause prun to exit.
93 Specifying Number of Processes
95 The following options specify the number of processes to launch. Note
96 that none of the options imply a particular binding policy - e.g., re‐
97 questing N processes for each socket does not imply that the processes
98 will be bound to the package.
99 Additional options and details are presented in prte-map(1). Below are
101 a few of the commonly used options.
102 -c, -n, --n, --np <#>
104 Run this many copies of the program on the given nodes. This
105 option indicates that the specified file is an executable pro‐
106 gram and not an application context. If no value is provided
107 for the number of copies to execute (i.e., neither the --np nor
108 its synonyms are provided on the command line), prun will auto‐
109 matically execute a copy of the program on each process slot
110 (see below for description of a “process slot”). This feature,
111 however, can only be used in the SPMD model and will return an
112 error (without beginning execution of the application) other‐
113 wise.
114 I/O Management
116 To manage standard I/O:
117 --output-filename <filename>
119 Redirect the stdout, stderr, and stddiag of all processes to a
120 process-unique version of the specified filename (“file‐
121 name.id”). Any directories in the filename will automatically
122 be created. Each output file will consist of “filename.id”,
123 where the id will be the processes’ rank, left-filled with ze‐
124 ro’s for correct ordering in listings. Both stdout and stderr
125 will be redirected to the file. A relative path value will be
126 converted to an absolute path based on the current working di‐
127 rectory where prun is executed. Note that this will not work in
128 environments where the file system on compute nodes differs from
129 that where prun is executed. This option accepts one case-in‐
130 sensitive directive, specified after a colon (:): NOCOPY indi‐
131 cates that the output is not to be echoed to the terminal.
132 --output-directory <path>
134 Redirect the stdout, stderr, and stddiag of all processes to a
135 process-unique location consisting of “//rank.id/std[out,err,di‐
136 ag]”, where the id will be the processes’ rank, left-filled with
137 zero’s for correct ordering in listings. Any directories in the
138 filename will automatically be created. A relative path value
139 will be converted to an absolute path based on the current work‐
140 ing directory where prun is executed. Note that this will not
141 work on environments where the file system on compute nodes dif‐
142 fers from that where prun is executed. This option also sup‐
143 ports two case-insensitive directives, specified in comma-delim‐
144 ited form after a colon (:): NOJOBID (omits the jobid directory
145 layer) and NOCOPY (do not copy the output to the terminal).
146 --stdin <rank>
148 The rank of the process that is to receive stdin. The default
149 is to forward stdin to rank 0, but this option can be used to
150 forward stdin to any process. It is also acceptable to specify
151 none, indicating that no processes are to receive stdin.
152 --merge-stderr-to-stdout
154 Merge stderr to stdout for each process.
155 --map-by :TAGOUTPUT
157 Tag each line of output to stdout, stderr, and stddiag with [jo‐
158 bid, MCW_rank]<stdxxx> indicating the jobid and rank of the
159 process that generated the output, and the channel which gener‐
160 ated it.
161 --map-by :TIMESTAMPOUTPUT
163 Timestamp each line of output to stdout, stderr, and stddiag.
164 --map-by :XMLOUTPUT
166 Provide all output to stdout, stderr, and stddiag in an xml for‐
167 mat.
168 --xterm <ranks>
170 Display the output from the processes identified by their ranks
171 in separate xterm windows. The ranks are specified as a comma-
172 separated list of ranges, with a -1 indicating all. A separate
173 window will be created for each specified process. Note: xterm
174 will normally terminate the window upon termination of the
175 process running within it. However, by adding a “!” to the end
176 of the list of specified ranks, the proper options will be pro‐
177 vided to ensure that xterm keeps the window open after the
178 process terminates, thus allowing you to see the process’ out‐
179 put. Each xterm window will subsequently need to be manually
180 closed. Note: In some environments, xterm may require that the
181 executable be in the user’s path, or be specified in absolute or
182 relative terms. Thus, it may be necessary to specify a local
183 executable as “./foo” instead of just “foo”. If xterm fails to
184 find the executable, prun will hang, but still respond correctly
185 to a ctrl-c. If this happens, please check that the executable
186 is being specified correctly and try again.
187 File and Environment Management
189 To manage files and runtime environment:
190 --path <path>
192 <path> that will be used when attempting to locate the requested
193 executables. This is used prior to using the local PATH set‐
194 ting.
195 --prefix <dir>
197 Prefix directory that will be used to set the PATH and LD_LI‐
198 BRARY_PATH on the remote node before invoking the target
199 process. See the “Remote Execution” section, below.
200 --noprefix
202 Disable the automatic --prefix behavior
203 -s | --preload-binary
205 Copy the specified executable(s) to remote machines prior to
206 starting remote processes. The executables will be copied to
207 the session directory and will be deleted upon completion of the
208 job.
209 --preload-files <files>
211 Preload the comma separated list of files to the current working
212 directory of the remote machines where processes will be
213 launched prior to starting those processes.
214 --set-cwd-to-session-dir
216 Set the working directory of the started processes to their ses‐
217 sion directory.
218 --wdir <dir>
220 Change to the directory <dir> before the user’s program exe‐
221 cutes. See the “Current Working Directory” section for notes on
222 relative paths. Note: If the --wdir option appears both on the
223 command line and in an application context, the context will
224 take precedence over the command line. Thus, if the path to the
225 desired working directory is different on the backend nodes,
226 then it must be specified as an absolute path that is correct
227 for the backend node.
228 --wd <dir>
230 Synonym for --wdir.
231 -x <env>
233 Export the specified environment variables to the remote nodes
234 before executing the program. Only one environment variable can
235 be specified per -x option. Existing environment variables can
236 be specified or new variable names specified with corresponding
237 values. If multiple -x options with the same variable name (re‐
238 gardless of value) are provided then the last one listed on the
239 command line will take precedence, and the others will be ig‐
240 nored. The exception to this is for PRTE_MCA_ prefixed environ‐
241 ment variables which will report an error in that scenario if
242 any of the values differ. For example: $ prun -x DISPLAY -x
243 OFILE=/tmp/out ...
244 The parser for the -x option is not very sophisticated; it does not
246 even understand quoted values. Users are advised to set variables in
247 the environment, and then use -x to export (not define) them.
248 MCA Parameters
250 Setting MCA parameters take a few different forms depending the target
251 project for the parameter. For example, MCA parameters targeting
252 OpenPMIx will contain the string pmix in their name, and MCA parameters
253 targeting PRTE will contain the string prte in their name. See the
254 “MCA” section, below, for finer details on the MCA.
255 --gpmixmca <key> <value>
257 Pass global PMIx MCA parameters that are applicable to all ap‐
258 plication contexts. <key> is the parameter name; <value> is the
259 parameter value.
260 --mca <key> <value>
262 Send arguments to various MCA modules. See the “MCA” section,
263 below.
264 --pmixmca <key> <value>
266 Send arguments to various PMIx MCA modules. See the “MCA” sec‐
267 tion, below.
268 --prtemca <key> <value>
270 Send arguments to various PRTE MCA modules. See the “MCA” sec‐
271 tion, below.
272 --pmixam <arg0>
274 Aggregate PMIx MCA parameter set file list. The arg0 argument
275 is a comma-separated list of tuning files. Each file containing
276 MCA parameter sets for this application context.
277 Debugging Options
279 --get-stack-traces
280 When paired with the --timeout option, prun will obtain and
281 print out stack traces from all launched processes that are
282 still alive when the timeout expires. Note that obtaining stack
283 traces can take a little time and produce a lot of output, espe‐
284 cially for large process-count jobs.
285 --timeout <seconds>
287 The maximum number of seconds that prun will run. After this
288 many seconds, prun will abort the launched job and exit with a
289 non-zero exit status. Using --timeout can be also useful when
290 combined with the --get-stack-traces option.
291 Other Options
293 There are also other options:
294 --allow-run-as-root
296 Allow prun to run when executed by the root user (prun defaults
297 to aborting when launched as the root user).
298 --app <appfile>
300 Provide an appfile, ignoring all other command line options.
301 --continuous
303 Job is to run until explicitly terminated.
304 --dvm-uri
306 Specify the URI of the DVM master, or the name of the file
307 (specified as file:filename) that contains that info.
308 --enable-recovery
310 Enable recovery from process failure [Default = disabled].
311 --disable-recovery
313 Disable recovery (resets all recovery options to off).
314 --map-by :DONOTLAUNCH
316 Perform all necessary operations to prepare to launch the appli‐
317 cation, but do not actually launch it.
318 --index-argv-by-rank
320 Uniquely index argv[0] for each process using its rank.
321 --max-restarts <num>
323 Max number of times to restart a failed process.
324 --pid PID of the daemon to which we should connect.
326 --report-child-jobs-separately
328 Return the exit status of the primary job only.
329 --show-progress
331 Output a brief periodic report on launch progress.
332 --terminate
334 Terminate the DVM.
335 The following options are useful for developers; they are not generally
337 useful to most users:
338 --map-by :DISPLAYALLOC
340 Display a detailed list of the allocation being used by this
341 job.
342 --map-by :DISPLAYDEVEL
344 Display a more detailed table showing the mapped location of
345 each process prior to launch.
346 --map-by :DISPLAYTOPO
348 Display the topology as part of the process map just before
349 launch.
350 --report-state-on-timeout
352 When paired with the --timeout command line option, report the
353 run-time subsystem state of each process when the timeout ex‐
354 pires.
355
357 One invocation of prun starts an application running under the PRTE
358 DVM. If the application is single process multiple data (SPMD), the
359 application can be specified on the prun command line.
360 If the application is multiple instruction multiple data (MIMD), com‐
362 prising of multiple programs, the set of programs and argument can be
363 specified in one of two ways: Extended Command Line Arguments, and Ap‐
364 plication Context.
365 An application context describes the MIMD program set including all ar‐
367 guments in a separate file. This file essentially contains multiple
368 prun command lines, less the command name itself. The ability to spec‐
369 ify different options for different instantiations of a program is an‐
370 other reason to use an application context.
371 Extended command line arguments allow for the description of the appli‐
373 cation layout on the command line using colons (:) to separate the
374 specification of programs and arguments. Some options are globally set
375 across all specified programs (e.g. --hostfile), while others are spe‐
376 cific to a single program (e.g. --np).
377 Specifying Host Nodes
379 Host nodes can be identified on the prun command line with the --host
380 option or in a hostfile. See prte-map(1) for more details.
381 Application Context or Executable Program?
383 To distinguish the two different forms, prun looks on the command line
384 for --app option. If it is specified, then the file named on the com‐
385 mand line is assumed to be an application context. If it is not speci‐
386 fied, then the file is assumed to be an executable program.
387 Locating Files
389 If no relative or absolute path is specified for a file, prun will
390 first look for files by searching the directories specified by the
391 --path option. If there is no --path option set or if the file is not
392 found at the --path location, then prun will search the user’s PATH en‐
393 vironment variable as defined on the source node(s).
394 If a relative directory is specified, it must be relative to the ini‐
396 tial working directory determined by the specific starter used. For
397 example when using the rsh or ssh starters, the initial directory is
398 $HOME by default. Other starters may set the initial directory to the
399 current working directory from the invocation of prun.
400 Current Working Directory
402 The --wdir prun option (and its synonym, --wd) allows the user to
403 change to an arbitrary directory before the program is invoked. It can
404 also be used in application context files to specify working directo‐
405 ries on specific nodes and/or for specific applications.
406 If the --wdir option appears both in a context file and on the command
408 line, the context file directory will override the command line value.
409 If the --wdir option is specified, prun will attempt to change to the
411 specified directory on all of the remote nodes. If this fails, prun
412 will abort.
413 If the --wdir option is not specified, prun will send the directory
415 name where prun was invoked to each of the remote nodes. The remote
416 nodes will try to change to that directory. If they are unable (e.g.,
417 if the directory does not exist on that node), then prun will use the
418 default directory determined by the starter.
419 All directory changing occurs before the user’s program is invoked.
421 Standard I/O
423 The PRTE DVM directs UNIX standard input to /dev/null on all processes
424 except the rank 0 process. The rank 0 process inherits standard input
425 from prun. Note: The node that invoked prun need not be the same as
426 the node where the rank 0 process resides. PRTE DVM handles the redi‐
427 rection of prun’s standard input to the rank 0 process.
428 The PRTE DVM directs UNIX standard output and error from remote nodes
430 to the node that invoked prun and prints it on the standard output/er‐
431 ror of prun. Local processes inherit the standard output/error of prun
432 and transfer to it directly.
433 Thus it is possible to redirect standard I/O for applications by using
435 the typical shell redirection procedure on prun.
436 $ prun --np 2 my_app < my_input > my_output
438 Note that in this example only the rank 0 process will receive the
440 stream from my_input on stdin. The stdin on all the other nodes will
441 be tied to /dev/null. However, the stdout from all nodes will be col‐
442 lected into the my_output file.
443 Signal Propagation
445 When prun receives a SIGTERM and SIGINT, it will attempt to kill the
446 entire job by sending all processes in the job a SIGTERM, waiting a
447 small number of seconds, then sending all processes in the job a
448 SIGKILL.
449 SIGUSR1 and SIGUSR2 signals received by prun are propagated to all pro‐
451 cesses in the job.
452 A SIGTSTOP signal to prun will cause a SIGSTOP signal to be sent to all
454 of the programs started by prun and likewise a SIGCONT signal to prun
455 will cause a SIGCONT sent.
456 Other signals are not currently propagated by prun.
458 Process Termination / Signal Handling
460 During the run of an application, if any process dies abnormally (ei‐
461 ther exiting before invoking PMIx_Finalize, or dying as the result of a
462 signal), prun will print out an error message and kill the rest of the
463 application.
464 Process Environment
466 Processes in the application inherit their environment from the PRTE
467 DVM daemon upon the node on which they are running. The environment is
468 typically inherited from the user’s shell. On remote nodes, the exact
469 environment is determined by the boot MCA module used. The rsh launch
470 module, for example, uses either rsh/ssh to launch the PRTE DVM daemon
471 on remote nodes, and typically executes one or more of the user’s
472 shell-setup files before launching the daemon. When running dynamical‐
473 ly linked applications which require the LD_LIBRARY_PATH environment
474 variable to be set, care must be taken to ensure that it is correctly
475 set when booting PRTE DVM.
476 See the “Remote Execution” section for more details.
478 Remote Execution
480 The PRTE DVM requires that the PATH environment variable be set to find
481 executables on remote nodes. This is typically only necessary in rsh-
482 or ssh-based environments. Batch and scheduled environments typically
483 copy the current environment to the execution of remote jobs, so if the
484 current environment has PATH and/or LD_LIBRARY_PATH set properly, the
485 remote nodes will also have it set properly. If the PRTE DVM was com‐
486 piled with shared library support, it may also be necessary to have the
487 LD_LIBRARY_PATH environment variable set on remote nodes as well (espe‐
488 cially to find the shared libraries required to run user applications).
489 However, it is not always desirable or possible to edit shell startup
491 files to set PATH and/or LD_LIBRARY_PATH. The --prefix option is pro‐
492 vided for some simple configurations where this is not possible.
493 The --prefix option takes a single argument: the base directory on the
495 remote node where PRTE DVM is installed. The PRTE DVM will use this
496 directory to set the remote PATH and LD_LIBRARY_PATH before executing
497 any user applications. This allows running jobs without having pre-
498 configured the PATH and LD_LIBRARY_PATH on the remote nodes.
499 The PRTE DVM adds the basename of the current node’s “bindir” (the di‐
501 rectory where the PRTE DVM’s executables are installed) to the prefix
502 and uses that to set the PATH on the remote node. Similarly, PRTE DVM
503 adds the basename of the current node’s “libdir” (the directory where
504 the PRTE DVM’s libraries are installed) to the prefix and uses that to
505 set the LD_LIBRARY_PATH on the remote node. For example:
506 Local bindir:
508 /local/node/directory/bin
509 Local libdir:
511 /local/node/directory/lib64
512 If the following command line is used:
514 $ prun --prefix /remote/node/directory
516 The PRTE DVM will add “/remote/node/directory/bin” to the PATH and
518 “/remote/node/directory/lib64” to the LD_LIBRARY_PATH on the remote
519 node before attempting to execute anything.
520 The --prefix option is not sufficient if the installation paths on the
522 remote node are different than the local node (e.g., if “/lib” is used
523 on the local node, but “/lib64” is used on the remote node), or if the
524 installation paths are something other than a subdirectory under a com‐
525 mon prefix.
526 Note that executing prun via an absolute pathname is equivalent to
528 specifying --prefix without the last subdirectory in the absolute path‐
529 name to prun.
530 For example:
532 $ /usr/local/bin/prun ...
534 is equivalent to
536 $ prun --prefix /usr/local ...
538 Exported Environment Variables
540 All environment variables that are named in the form PMIX_\* will auto‐
541 matically be exported to new processes on the local and remote nodes.
542 Environmental parameters can also be set/forwarded to the new processes
543 using the MCA parameter mca_base_env_list. While the syntax of the -x
544 option and MCA param allows the definition of new variables, note that
545 the parser for these options are currently not very sophisticated - it
546 does not even understand quoted values. Users are advised to set vari‐
547 ables in the environment and use the option to export them; not to de‐
548 fine them.
549 Setting MCA Parameters
551 The --mca / --pmixmca / --prtemca switches (referenced here as “--mca
552 switches” for brevity) allow the passing of parameters to various MCA
553 (Modular Component Architecture) modules. MCA modules have direct im‐
554 pact on programs because they allow tunable parameters to be set at run
555 time.
556 The -mca switch takes two arguments: <key> and <value>. The <key> ar‐
558 gument generally specifies which MCA module will receive the value.
559 For example, the <key> “rmaps” is used to select which RMAPS to be used
560 for mapping processes to nodes. The <value> argument is the value that
561 is passed. For example:
562 prun -prtemca rmaps seq -np 1 foo
564 Tells PRTE to use the “seq” RMAPS component, and to run a single
565 copy of “a.out” on an allocated node.
566 The -mca switch can be used multiple times to specify different <key>
568 and/or <value> arguments. If the same <key> is specified more than
569 once, the <value>s are concatenated with a comma (“,”) separating them.
570 Note that the -mca switch is simply a shortcut for setting environment
572 variables. The same effect may be accomplished by setting correspond‐
573 ing environment variables before running prun. The form of the envi‐
574 ronment variables depends on the type of the --mca switch.
575 --mca PRTE_MCA_<key>=<value>
577 --pmixmca
579 PMIX_MCA_<key>=<value>
580 --prtemca
582 `PRTE_MCA_=``
583 Thus, the -mca switch overrides any previously set environment vari‐
585 ables. The -mca settings similarly override MCA parameters set in the
586 $PRTE_PREFIX/etc/prte-mca-params.conf or $HOME/.prte/mca-params.conf
587 file.
588 Unknown <key> arguments are still set as environment variable – they
590 are not checked (by prun) for correctness. Illegal or incorrect <val‐
591 ue> arguments may or may not be reported – it depends on the specific
592 MCA module.
593 To find the available component types under the MCA architecture, or to
595 find the available parameters for a specific component, use the pinfo
596 command. See the pinfo(1) man page for detailed information on the
597 command.
598 Running as root
600 The PRTE team strongly advises against executing prun as the root user.
601 Applications should be run as regular (non-root) users.
602 Reflecting this advice, prun will refuse to run as root by default. To
604 override this default, you can add the --allow-run-as-root option to
605 the prun command line.
606
608 There is no standard definition for what prun should return as an exit
609 status. After considerable discussion, we settled on the following
610 method for assigning the prun exit status (note: in the following de‐
611 scription, the “primary” job is the initial application started by prun
612 - all jobs that are spawned by that job are designated “secondary”
613 jobs):
614 • if all processes in the primary job normally terminate with exit sta‐
616 tus 0, we return 0
617 • if one or more processes in the primary job normally terminate with
619 non-zero exit status, we return the exit status of the process with
620 the lowest rank to have a non-zero status
621 • if all processes in the primary job normally terminate with exit sta‐
623 tus 0, and one or more processes in a secondary job normally termi‐
624 nate with non-zero exit status, we (a) return the exit status of the
625 process with the lowest rank in the lowest jobid to have a non-zero
626 status, and (b) output a message summarizing the exit status of the
627 primary and all secondary jobs.
628 • if the cmd line option --report-child-jobs-separately is set, we will
630 return -only- the exit status of the primary job. Any non-zero exit
631 status in secondary jobs will be reported solely in a summary print
632 statement.
633 By default, the job will abort when any process terminates with non-ze‐
635 ro status. The MCA parameter prte_abort_on_non_zero_status can be set
636 to false (or 0) to cause the PRTE DVM to not abort a job if one or more
637 processes return a non-zero status. In that situation the PRTE DVM
638 records and notes that processes exited with non-zero termination sta‐
639 tus to report the approprate exit status of prun (per bullet points
640 above).
641 If the --timeout command line option is used and the timeout expires
643 before the job completes (thereby forcing prun to kill the job) prun
644 will return an exit status equivalent to the value of ETIMEDOUT (which
645 is typically 110 on Linux and OS X systems).
6462021-08-23 prun(1)