mpprun

Name

mpprun -- Run a series of commands on a Scyld cluster using a dynamically generated job map.

Synopsis

mpprun [-h, --help] [-V, --version] [-p,--prefix] command [command-args...] [--map node list] [--all-cpus]
[--all-nodes] [--np processes] [--all-local] [--no-local] [--exclude node list]

Description

The mpprun program sequentially runs the specified program on a dynamically selected set of cluster nodes. It generates a job map from the currently installed beomap scheduler, and runs the program on each node specified in the map. The scheduling parameters from the command line and environment are the same as for beomap, and the resulting job map is identical to the job map that beomap would if generate at that instant in time for that program name.

mpprun is similar to the beorun program, but beorun starts the job simultaneously on the cluster nodes.

Options

The following general command line options are available to mpprun. Also see the next section, which describes the beomap job map parameters.

-h, --help, -u, --usage

Print the command usage message on stdout and exit. When one of these options is recognized in the option list, all following options will be ignored.

-V

Print the command version number on stdout and exit. Any following options will be ignored.

-p,--prefix

Prefix each line of output the node number.

Unrecognized options and invalid option formats are reported on stderr and the command exits with exit status 1 (invalid option) or 2 (no command specified or invalid command).

Job Map Parameters

You can influence the beomap job map either by entering command line options or by setting environment variables. Following are the available command line options, with their equivalent environment variables; note that the command line options take precedence over the environment variables.

All of the beomap job map parameters listed below can also be used directly with beorun, mpirun, and mpprun.

--map node list

Specify a process map consisting of a colon-delimited list of nodes. Each node in the list indicates where one process will be assigned. The number of entries in the job map implies the number of ranks in the job.

Listing a node more than once in the list will assign multiple processes to that node. Typically, this is done to assign one process to each processor or core on a node, but this can also be used to assign more processes to a node than it has processors or cores.

The --all-cpus, --all-nodes, --np processes, --all-local, --no-local, and --exclude parameters are not specified with the --map parameter.

The equivalent environment variable is BEOWULF_JOB_MAP=node list.

--all-cpus

Create a process map consisting of all "up" nodes, with each node number repeated to represent the number of CPUs on that node. This parameter is not specified with the --map parameter.

The equivalent environment variable is ALL_CPUS.

--all-nodes

Create a process map consisting of all "up" nodes, with 1 CPU mapped on each of the "up" nodes. This parameter is not specified with the --map parameter.

The equivalent environment variable is ALL_NODES.

NOTE: --all-nodes and ALL_NODES do not currently work as intended; this is a known issue in ClusterWare 4.1.0. Refer to the Release Notes for suggested workarounds.

--np processes

Specify the number of processes to run. The beomap command attempts to place one process per processor or core, but will assign multiple processes per processor or core if there are not enough individual processors or cores available. This parameter is not specified with the --map parameter.

The equivalent environment variable is NP=processes.

--all-local

Create a process map consisting entirely of master node entries. This option eliminates everything except node -1 from the pool of candidate node numbers, thus forcing the map to use node -1 for everything. This parameter is not specified with the --map parameter.

This option is a handy shortcut for troubleshooting connectivity problems or testing on an isolated head node before running a job on a "live" cluster.

The equivalent environment variable is ALL_LOCAL.

--no-local

Exclude the master in the process map. This option is essentially a syntactic shortcut for including -1 in the --exclude node list option. For MPI jobs, this option puts the "rank 0" job on a compute node instead of the master. This parameter is not specified with the --map parameter.

The equivalent environment variable is NO_LOCAL.

--exclude node list

Do not include the listed nodes in the process map; the nodes to be excluded are stated as a colon-delimited list. This parameter is not specified with the --map parameter.

The equivalent environment variable is EXCLUDE=node list.

Tip

The environment variables have an order of priority. The BEOWULF_JOB_MAP variable acts as a "master override" for the other environment variables. If BEOWULF_JOB_MAP is not set, the following priorities apply:

  • Three of the environment variables determine how many ranks to schedule in the map: ALL_CPUS (priority 1), ALL_NODES (priority 2), and NP (priority 3). If none of these are set explicitly by the user, then a usage of NP=1 is assumed.

  • Three of the environment variables determine what node numbers are candidates for being mapped: ALL_LOCAL (priority 1), NO_LOCAL (priority 2), and EXCLUDE (priority 3).

Caution

It is an error to use NO_LOCAL and ALL_LOCAL together. If both are used, ALL_LOCAL will take precedence.

Examples

Run uptime on any two available cluster compute nodes.

[user@cluster user] $ mpprun --np 2 --no-local uptime
  11:05am  up 2 days, 11:16,  0 users,  load average: 0.05, 0.24, 0.65
  11:05am  up 2 days, 11:16,  0 users,  load average: 0.01, 0.07, 0.37

See Also

beomap(1), bpsh(1), mpprun(1), beonpc(1), and the User's Guide.