1MPI_Comm_spawn_multiple(3) LAM/MPI MPI_Comm_spawn_multiple(3)
2
6 MPI_Comm_spawn_multiple - Spawn a dynamic MPI process from multiple
7 executables
8
10 #include <mpi.h>
11 int
12 MPI_Comm_spawn_multiple(int count, char **commands, char ***argvs,
13 int *maxprocs, MPI_Info *infos, int root,
14 MPI_Comm comm, MPI_Comm *intercomm,
15 int *errcodes)
16
18 count - number of commands (only significant at root)
19 commands
20 - commands to be executed (only significant at root)
21 argvs - arguments for commands (only significant at root)
22 maxprocs
23 - max number of processes for each command (only significant at
24 root)
25 infos - startup hints for each command
26 root - rank of process to perform the spawn
27 comm - parent intracommunicator
28
31 intercomm
32 - child intercommunicator containing spawned processes
33 errcodes
34 - one code per process
35
38 A group of processes can create another group of processes with
39 MPI_Comm_spawn_multiple . This function is a collective operation over
40 the parent communicator. The child group starts up like any MPI appli‐
41 cation. The processes must begin by calling MPI_Init , after which the
42 pre-defined communicator, MPI_COMM_WORLD , may be used. This world
43 communicator contains only the child processes. It is distinct from
44 the MPI_COMM_WORLD of the parent processes.
45 MPI_Comm_spawn_multiple is used to manually specify a group of differ‐
47 ent executables and arguments to spawn. MPI_Comm_spawn is used to
48 specify one executable and set of arguments (although a LAM/MPI app‐
49 schema(5) can be provided to MPI_Comm_spawn via the "file" info key).
50 Communication With Spawned Processes
52 The natural communication mechanism between two groups is the intercom‐
54 municator. The second communicator argument to MPI_Comm_spawn_multiple
55 returns an intercommunicator whose local group contains the parent pro‐
56 cesses (same as the first communicator argument) and whose remote group
57 contains child processes. The child processes can access the same
58 intercommunicator by using the MPI_Comm_get_parent call. The remote
59 group size of the parent communicator is zero if the process was cre‐
60 ated by mpirun (1) instead of one of the spawn functions. Both groups
61 can decide to merge the intercommunicator into an intracommunicator
62 (with the MPI_Intercomm_merge () function) and take advantage of other
63 MPI collective operations. They can then use the merged intracommuni‐
64 cator to create new communicators and reach other processes in the MPI
65 application.
66 Resource Allocation
68 Note that no MPI_Info keys are recognized by this implementation of
70 MPI_Comm_spawn_multiple . To use the "file" info key to specify an
71 appschema(5), use LAM's MPI_Comm_spawn . This may be preferable to
72 MPI_Comm_spawn_multiple because it allows the arbitrary specification
73 of what nodes and/or CPUs should be used to launch jobs (either SPMD or
74 MPMD). See MPI_Comm_spawn(3) for more details.
75 The value of MPI_INFO_NULL should be given for each value in infos (the
77 infos array is not currently examined by LAM/MPI, so specifying non-
78 NULL values for the array values is not harmful). LAM schedules the
79 given number of processes onto LAM nodes by starting with CPU 0 (or the
80 lowest numbered CPU), and continuing through higher CPU numbers, plac‐
81 ing one process on each CPU. If the process count is greater than the
82 CPU count, the procedure repeats.
83 Process Terminiation
85 Note that the process[es] spawned by MPI_COMM_SPAWN (and
87 MPI_COMM_SPAWN_MULTIPLE ) effectively become orphans. That is, the
88 spawnning MPI application does not wait for the spawned application to
89 finish. Hence, there is no guarantee the spawned application has fin‐
90 ished when the spawning completes. Similarly, killing the spawning
91 application will also have no effect on the spawned application.
92 User applications can effect this kind of behavior with MPI_BARRIER
94 between the spawning and spawned processed before MPI_FINALIZE .
95 Note that lamclean will kill *all* MPI processes.
98 Process Count
100 The maxprocs array parameter to MPI_Comm_spawn_multiple specifies the
102 exact number of processes to be started. If it is not possible to
103 start the desired number of processes, MPI_Comm_spawn_multiple will
104 return an error code. Note that even though maxprocs is only relevant
105 on the root, all ranks must have an errcodes array long enough to han‐
106 dle an integer error code for every process that tries to launch, or
107 give MPI constant MPI_ERRCODES_IGNORE for the errcodes argument. While
108 this appears to be a contradiction, it is per the MPI-2 standard. :-\
109 Frequently, an application wishes to chooses a process count so as to
111 fill all processors available to a job. MPI indicates the maximum num‐
112 ber of processes recommended for a job in the pre-defined attribute,
113 MPI_UNIVERSE_SIZE , which is cached on MPI_COMM_WORLD .
114 The typical usage is to subtract the value of MPI_UNIVERSE_SIZE from
116 the number of processes currently in the job and spawn the difference.
117 LAM sets MPI_UNIVERSE_SIZE to the number of CPUs in the user's LAM ses‐
118 sion (as defined in the boot schema [bhost(5)] via lamboot (1)).
119 See MPI_Init(3) for other pre-defined attributes that are helpful when
121 spawning.
122 Locating an Executable Program
124 The executable program file must be located on the node(s) where the
126 process(es) will run. On any node, the directories specified by the
127 user's PATH environment variable are searched to find the program.
128 All MPI runtime options selected by mpirun (1) in the initial applica‐
130 tion launch remain in effect for all child processes created by the
131 spawn functions.
132 Command-line Arguments
134 The argvs array parameter to MPI_Comm_spawn_multiple should not contain
136 the program name since it is given in the first parameter. The command
137 line that is passed to the newly launched program will be the program
138 name followed by the strings in corresponding entry in the argvs array.
139
142 The IMPI standard only supports MPI-1 functions. Hence, this function
143 is currently not designed to operate within an IMPI job.
144
147 If an error occurs in an MPI function, the current MPI error handler is
148 called to handle it. By default, this error handler aborts the MPI
149 job. The error handler may be changed with MPI_Errhandler_set ; the
150 predefined error handler MPI_ERRORS_RETURN may be used to cause error
151 values to be returned (in C and Fortran; this error handler is less
152 useful in with the C++ MPI bindings. The predefined error handler
153 MPI::ERRORS_THROW_EXCEPTIONS should be used in C++ if the error value
154 needs to be recovered). Note that MPI does not guarantee that an MPI
155 program can continue past an error.
156 All MPI routines (except MPI_Wtime and MPI_Wtick ) return an error
158 value; C routines as the value of the function and Fortran routines in
159 the last argument. The C++ bindings for MPI do not return error val‐
160 ues; instead, error values are communicated by throwing exceptions of
161 type MPI::Exception (but not by default). Exceptions are only thrown
162 if the error value is not MPI::SUCCESS .
163 Note that if the MPI::ERRORS_RETURN handler is set in C++, while MPI
166 functions will return upon an error, there will be no way to recover
167 what the actual error value was.
168 MPI_SUCCESS
169 - No error; MPI routine completed successfully.
170 MPI_ERR_COMM
171 - Invalid communicator. A common error is to use a null commu‐
172 nicator in a call (not even allowed in MPI_Comm_rank ).
173 MPI_ERR_SPAWN
174 - Spawn error; one or more of the applications attempting to be
175 launched failed. Check the returned error code array.
176 MPI_ERR_ARG
177 - Invalid argument. Some argument is invalid and is not identi‐
178 fied by a specific error class. This is typically a NULL
179 pointer or other such error.
180 MPI_ERR_ROOT
181 - Invalid root. The root must be specified as a rank in the
182 communicator. Ranks must be between zero and the size of the
183 communicator minus one.
184 MPI_ERR_OTHER
185 - Other error; use MPI_Error_string to get more information
186 about this error code.
187 MPI_ERR_INTERN
188 - An internal error has been detected. This is fatal. Please
189 send a bug report to the LAM mailing list (see http://www.lam-
190 mpi.org/contact.php ).
191
194 appschema(5), bhost(5), lamboot(1), MPI_Comm_get_parent(3), MPI_Inter‐
195 comm_merge(3), MPI_Comm_spawn_multiple(3), MPI_Info_create(3),
196 MPI_Info_set(3), MPI_Info_delete(3), MPI_Info_free(3), MPI_Init(3),
197 mpirun(1)
198
201 For more information, please see the official MPI Forum web site, which
202 contains the text of both the MPI-1 and MPI-2 standards. These docu‐
203 ments contain detailed information about each MPI function (most of
204 which is not duplicated in these man pages).
205 http://www.mpi-forum.org/
207
211 The LAM Team would like the thank the MPICH Team for the handy program
212 to generate man pages ("doctext" from ftp://ftp.mcs.anl.gov/pub/sow‐
213 ing/sowing.tar.gz ), the initial formatting, and some initial text for
214 most of the MPI-1 man pages.
215
217 spawnmult.c
218LAM/MPI 7.1.2 2/23/2006 MPI_Comm_spawn_multiple(3)