Location Nomenclature¶

As described above, mpirun can specify arbitrary locations in the current LAM universe. Locations can be specified either by CPU or by node (noted by the "where" in the SYNTAX section, above). Note that LAM does not bind processes to CPUs -- specifying a location "by CPU" is really a convenience mechanism for SMPs that ultimately maps down to a specific node.

Note that LAM effectively numbers MPI_COMM_WORLD ranks from left-to-right in the where, regardless of which nomenclature is used. This can be important because typical MPI programs tend to communicate more with their immediate neighbors (i.e., myrank +/- X) than distant neighbors. When neighbors end up on the same node, the shmem RPIs can be used for communication rather than the network RPIs, which can result in faster MPI performance.

Specifying locations by node will launch one copy of an executable per specified node. Using a capitol "N" tells LAM to use all available nodes that were lambooted (see lamboot(1)). Ranges of specific nodes can also be specified in the form "nR[,R]*", where R specifies either a single node number or a valid range of node numbers in the range of [0, num_nodes). For example:

mpirun N a.out

Runs one copy of the the executable a.out on all available nodes in the LAM universe. MPI_COMM_WORLD rank 0 will be on n0, rank 1 will be on n1, etc.

mpirun n0-3 a.out

Runs one copy of the the executable a.out on nodes 0 through 3. MPI_COMM_WORLD rank 0 will be on n0, rank 1 will be on n1, etc.

mpirun n0-3,8-11,15 a.out

Runs one copy of the the executable a.out on nodes 0 through 3, 8 through 11, and 15. MPI_COMM_WORLD ranks will be ordered as follows: (0, n0), (1, n1), (2, n2), (3, n3), (4, n8), (5, n9), (6, n10), (7, n11), (8, n15).

Specifying by CPU is the preferred method of launching MPI jobs. The intent is that the boot schema used with lamboot(1) will indicate how many CPUs are available on each node, and then a single, simple mpirun command can be used to launch across all of them. As noted above, specifying CPUs does not actually bind processes to CPUs -- it is only a convenience mechanism for launching on SMPs. Otherwise, the by-CPU notation is the same as the by-node notation, except that "C" and "c" are used instead of "N" and "n".

Assume in the following example that the LAM universe consists of four 4-way SMPs. So c0-3 are on n0, c4-7 are on n1, c8-11 are on n2, and 13-15 are on n3.

mpirun C a.out

Runs one copy of the the executable a.out on all available CPUs in the LAM universe. This is typically the simplest (and preferred) method of launching all MPI jobs (even if it resolves to one process per node). MPI_COMM_WORLD ranks 0-3 will be on n0, ranks 4-7 will be on n1, ranks 8-11 will be on n2, and ranks 13-15 will be on n3.

mpirun c0-3 a.out

Runs one copy of the the executable a.out on CPUs 0 through 3. All four ranks of MPI_COMM_WORLD will be on MPI_COMM_WORLD.

mpirun c0-3,8-11,15 a.out

Runs one copy of the the executable a.out on CPUs 0 through 3, 8 through 11, and 15. MPI_COMM_WORLD ranks 0-3 will be on n0, 4-7 will be on n2, and 8 will be on n3.

The reason that the by-CPU nomenclature is preferred over the by-node nomenclature is best shown through example. Consider trying to run the first CPU example (with the same MPI_COMM_WORLD mapping) with the by-node nomenclature -- run one copy of a.out for every available CPU, and maximize the number of local neighbors to potentially maximize MPI performance. One solution would be to use the following command:

mpirun n0,0,0,0,1,1,1,1,2,2,2,2,3,3,3,3 a.out

This works, but is definitely klunky to type. It is typically easier to use the by-CPU notation. One might think that the following is equivalent:

mpirun N -np 16 a.out

This is not equivalent because the MPI_COMM_WORLD rank mappings will be assigned by node rather than by CPU. Hence rank 0 will be on n0, rank 1 will be on n1, etc. Note that the following, however, is equivalent, because LAM interprets lack of a where as "C":

mpirun -np 16 a.out

However, a "C" can tend to be more convenient, especially for batch-queuing scripts because the exact number of processes may vary between queue submissions. Since the batch system will determine the final number of CPUs available, having a generic script that effectively says "run on everything you gave me" may lead to more portable / re-usable scripts.

Finally, it should be noted that specifying multiple where clauses are perfectly acceptable. As such, mixing of the by-node and by-CPU syntax is also valid, albiet typically not useful. For example:

mpirun C N a.out

However, in some cases, specifying multiple where clauses can be useful. Consider a parallel application where MPI_COMM_WORLD rank 0 will be a "manager" and therefore consume very few CPU cycles because it is usually waiting for "worker" processes to return results. Hence, it is probably desirable to run one "worker" process on all available CPUs, and run one extra process that will be the "manager":

mpirun c0 C manager-worker-program

Application Schema or Executable Program?¶

To distinguish the two different forms, mpirun looks on the command line for where or the -c option. If neither is specified, then the file named on the command line is assumed to be an application schema. If either one or both are specified, then the file is assumed to be an executable program. If where and -c both are specified, then copies of the program are started on the specified nodes/CPUs according to an internal LAM scheduling policy. Specifying just one node effectively forces LAM to run all copies of the program in one place. If -c is given, but not where, then all available CPUs on all LAM nodes are used. If where is given, but not -c, then one copy of the program is run on each node.

Program Transfer¶

By default, LAM searches for executable programs on the target node where a particular instantiation will run. If the file system is not shared, the target nodes are homogeneous, and the program is frequently recompiled, it can be convenient to have LAM transfer the program from a source node (usually the local node) to each target node. The -s option specifies this behavior and identifies the single source node.

Locating Files¶

LAM looks for an executable program by searching the directories in the user's PATH environment variable as defined on the source node(s). This behavior is consistent with logging into the source node and executing the program from the shell. On remote nodes, the "." path is the home directory.

LAM looks for an application schema in three directories: the local directory, the value of the LAMAPPLDIR environment variable, and laminstalldir/boot, where "laminstalldir" is the directory where LAM/MPI was installed.

Standard I/O¶

LAM directs UNIX standard input to /dev/null on all remote nodes. On the local node that invoked mpirun, standard input is inherited from mpirun. The default is what used to be the -w option to prevent conflicting access to the terminal.

LAM directs UNIX standard output and error to the LAM daemon on all remote nodes. LAM ships all captured output/error to the node that invoked mpirun and prints it on the standard output/error of mpirun. Local processes inherit the standard output/error of mpirun and transfer to it directly.

Thus it is possible to redirect standard I/O for LAM applications by using the typical shell redirection procedure on mpirun.

Note that in this example only the local node (i.e., the node where mpirun was invoked from) will receive the stream from my_input on stdin. The stdin on all the other nodes will be tied to /dev/null. However, the stdout from all nodes will be collected into the my_output file.

The -f option avoids all the setup required to support standard I/O described above. Remote processes are completely directed to /dev/null and local processes inherit file descriptors from lamboot(1).

Pseudo-tty support¶

The -pty option enabled pseudo-tty support for process output (it is also enabled by default). This allows, among other things, for line buffered output from remote nodes (which is probably what you want). This option can be disabled with the -npty switch.

Process Termination / Signal Handling¶

During the run of an MPI application, if any rank dies abnormally (either exiting before invoking MPI_FINALIZE, or dying as the result of a signal), mpirun will print out an error message and kill the rest of the MPI application.

By default, LAM/MPI only installs a signal handler for one signal in user programs (SIGUSR2 by default, but this can be overridden when LAM is configured and built). Therefore, it is safe for users to install their own signal handlers in LAM/MPI programs (LAM notices death-by-signal cases by examining the process' return status provided by the operating system).

User signal handlers should probably avoid trying to cleanup MPI state -- LAM is neither thread-safe nor async-signal-safe. For example, if a seg fault occurs in MPI_SEND (perhaps because a bad buffer was passed in) and a user signal handler is invoked, if this user handler attempts to invoke MPI_FINALIZE, Bad Things could happen since LAM/MPI was already "in" MPI when the error occurred. Since mpirun will notice that the process died due to a signal, it is probably not necessary (and safest) for the user to only clean up non-MPI state.

If the -sigs option is used with mpirun, LAM/MPI will install several signal handlers to locally on each rank to catch signals, print out error messages, and kill the rest of the MPI application. This is somewhat redundant behavior since this is now all handled by mpirun, but it has been left for backwards compatability.

Process Exit Statuses¶

The -sa,   -sf, and -p parameters can be used to display the exist statuses of the individual MPI processes as they terminate. -sa forces the exit statuses to be displayed for all processes; -sf only displays the exist statuses if at least one process terminates either by a signal or a non-zero exit status (note that exiting before invoking MPI_FINALIZE will cause a non-zero exit status).

The status of each process is printed out, one per line, in the following format:

If killed is 1, then status is the signal number. If killed is 0, then status is the exit status of the process.

The default prefix_string is "mpirun:", but the -p option can be used override this string.

Current Working Directory¶

The default behavior of mpirun has changed with respect to the directory that processes will be started in.

The -wd option to mpirun allows the user to change to an arbitrary directory before their program is invoked. It can also be used in application schema files to specify working directories on specific nodes and/or for specific applications.

If the -wd option appears both in a schema file and on the command line, the schema file directory will override the command line value.

The -D option will change the current working directory to the directory where the executable resides. It cannot be used in application schema files. -wd is mutually exclusive with -D.

If neither -wd nor -D are specified, the local node will send the directory name where mpirun was invoked from to each of the remote nodes. The remote nodes will then try to change to that directory. If they fail (e.g., if the directory does not exists on that node), they will start with from the user's home directory.

All directory changing occurs before the user's program is invoked; it does not wait until MPI_INIT is called.

Process Environment¶

Processes in the MPI application inherit their environment from the LAM daemon upon the node on which they are running. The environment of a LAM daemon is fixed upon booting of the LAM with lamboot(1) and is typically inherited from the user's shell. On the origin node, this will be the shell from which lamboot(1) was invoked; on remote nodes, the exact environment is determined by the boot SSI module used by lamboot(1). The rsh boot module, for example, uses either rsh/ssh to launch the LAM daemon on remote nodes, and typically executes one or more of the user's shell-setup files before launching the LAM daemon. When running dynamically linked applications which require the LD_LIBRARY_PATH environment variable to be set, care must be taken to ensure that it is correctly set when booting the LAM.

Exported Environment Variables¶

All environment variables that are named in the form LAM_MPI_*, LAM_IMPI_*, or IMPI_* will automatically be exported to new processes on the local and remote nodes. This exporting may be inhibited with the -nx option.

Additionally, the -x option to mpirun can be used to export specific environment variables to the new processes. While the syntax of the -x option allows the definition of new variables, note that the parser for this option is currently not very sophisticated - it does not even understand quoted values. Users are advised to set variables in the environment and use -x to export them; not to define them.

Trace Generation¶

Two switches control trace generation from processes running under LAM and both must be in the on position for traces to actually be generated. The first switch is controlled by mpirun and the second switch is initially set by mpirun but can be toggled at runtime with MPIL_Trace_on(2) and MPIL_Trace_off(2). The -t (-ton is equivalent) and -toff options all turn on the first switch. Otherwise the first switch is off and calls to MPIL_Trace_on(2) in the application program are ineffective. The -t option also turns on the second switch. The -toff option turns off the second switch. See MPIL_Trace_on(2) and lamtrace(1) for more details.

MPI Data Conversion¶

LAM's MPI library converts MPI messages from local representation to LAM representation upon sending them and then back to local representation upon receiving them. If the case of a LAM consisting of a homogeneous network of machines where the local representation differs from the LAM representation this can result in unnecessary conversions.

The -O switch used to be necessary to indicate to LAM whether the mulitcomputer was homogeneous or not. LAM now automatically determines whether a given MPI job is homogeneous or not. The -O flag will silently be accepted for backwards compatability, but it is ignored.

SSI (System Services Interface)¶

The -ssi switch allows the passing of parameters to various SSI modules. LAM's SSI modules are described in detail in lamssi(7). SSI modules have direct impact on MPI programs because they allow tunable parameters to be set at run time (such as which RPI communication device driver to use, what parameters to pass to that RPI, etc.).

The -ssi switch takes two arguments: key and value. The key argument generally specifies which SSI module will receive the value. For example, the key "rpi" is used to select which RPI to be used for transporting MPI messages. The value argument is the value that is passed. For example:

mpirun -ssi rpi lamd N foo

Tells LAM to use the "lamd" RPI and to run a single copy of "foo" on every node.

mpirun -ssi rpi tcp N foo

Tells LAM to use the "tcp" RPI.

mpirun -ssi rpi sysv N foo

Tells LAM to use the "sysv" RPI.

And so on. LAM's RPI SSI modules are described in lamssi_rpi(7).

The -ssi switch can be used multiple times to specify different key and/or value arguments. If the same key is specified more than once, the values are concatenated with a comma (",") separating them.

Note that the -ssi switch is simply a shortcut for setting environment variables. The same effect may be accomplished by setting corresponding environment variables before running mpirun. The form of the environment variables that LAM sets are: LAM_MPI_SSI_key=value.

Note that the -ssi switch overrides any previously set environment variables. Also note that unknown key arguments are still set as environment variable -- they are not checked (by mpirun) for correctness. Illegal or incorrect value arguments may or may not be reported -- it depends on the specific SSI module.

The -ssi switch obsoletes the old -c2c and -lamd switches. These switches used to be relevant because LAM could only have two RPI's available at a time: the lamd RPI and one of the C2C RPIs. This is no longer true -- all RPI's are now available and choosable at run-time. Selecting the lamd RPI is shown in the examples above. The -c2c switch has no direct translation since "C2C" used to refer to all other RPI's that were not the lamd RPI. As such, -ssi rpi value must be used to select the specific desired RPI (whether it is "lamd" or one of the other RPI's).

Guaranteed Envelope Resources¶

By default, LAM will guarantee a minimum amount of message envelope buffering to each MPI process pair and will impede or report an error to a process that attempts to overflow this system resource. This robustness and debugging feature is implemented in a machine specific manner when direct communication is used. For normal LAM communication via the LAM daemon, a protocol is used. The -nger option disables GER and the measures taken to support it. The minimum GER is configured by the system administrator when LAM is installed. See MPI(7) for more details.