Slots are not hardware resources

Slots are frequently incorrectly conflated with hardware resources. It is important to realize that slots are an entirely different metric than the number (and type) of hardware resources available.

Here are some examples that may help illustrate the difference:

1. More processor cores than slots

Consider a resource manager job environment that tells Open MPI that there is a single node with 20 processor cores and 2 slots available. By default, Open MPI will only let you run up to 2 processes.

Meaning: you run out of slots long before you run out of processor cores.

2. More slots than processor cores

Consider a hostfile with a single node listed with a "slots=50" qualification. The node has 20 processor cores. By default, Open MPI will let you run up to 50 processes.

Meaning: you can run many more processes than you have processor cores.

Environment Variables

MPIEXEC_TIMEOUT

Synonym for the --timeout command line option.

Specifying Host Nodes

Host nodes can be identified on the mpirun command line with the -host option or in a hostfile.

For example,

mpirun -H aa,aa,bb ./a.out

launches two processes on node aa and one on bb.

Or, consider the hostfile

% cat myhostfile
aa slots=2
bb slots=2
cc slots=2

Here, we list both the host names (aa, bb, and cc) but also how many slots there are for each.

mpirun -hostfile myhostfile ./a.out

will launch two processes on each of the three nodes.

mpirun -hostfile myhostfile -host aa ./a.out

will launch two processes, both on node aa.

mpirun -hostfile myhostfile -host dd ./a.out

will find no hosts to run on and abort with an error. That is, the specified host dd is not in the specified hostfile.

When running under resource managers (e.g., SLURM, Torque, etc.), Open MPI will obtain both the hostnames and the number of slots directly from the resource manger.

Specifying Number of Processes

As we have just seen, the number of processes to run can be set using the hostfile. Other mechanisms exist.

The number of processes launched can be specified as a multiple of the number of nodes or processor sockets available. For example,

mpirun -H aa,bb -npersocket 2 ./a.out

launches processes 0-3 on node aa and process 4-7 on node bb, where aa and bb are both dual-socket nodes. The -npersocket option also turns on the -bind-to-socket option, which is discussed in a later section.

mpirun -H aa,bb -npernode 2 ./a.out

launches processes 0-1 on node aa and processes 2-3 on node bb.

mpirun -H aa,bb -npernode 1 ./a.out

launches one process per host node.

mpirun -H aa,bb -pernode ./a.out

is the same as -npernode 1.

Another alternative is to specify the number of processes with the -np option. Consider now the hostfile

% cat myhostfile
aa slots=4
bb slots=4
cc slots=4

Now,

mpirun -hostfile myhostfile -np 6 ./a.out

will launch processes 0-3 on node aa and processes 4-5 on node bb. The remaining slots in the hostfile will not be used since the -np option indicated that only 6 processes should be launched.

Mapping Processes to Nodes: Using Policies

The examples above illustrate the default mapping of process processes to nodes. This mapping can also be controlled with various mpirun options that describe mapping policies.

Consider the same hostfile as above, again with -np 6:

node aa node bb node cc

mpirun 0 1 2 3 4 5

mpirun --map-by node 0 3 1 4 2 5

mpirun -nolocal 0 1 2 3 4 5

The --map-by node option will load balance the processes across the available nodes, numbering each process in a round-robin fashion.

The -nolocal option prevents any processes from being mapped onto the local host (in this case node aa). While mpirun typically consumes few system resources, -nolocal can be helpful for launching very large jobs where mpirun may actually need to use noticeable amounts of memory and/or processing time.

Just as -np can specify fewer processes than there are slots, it can also oversubscribe the slots. For example, with the same hostfile:

mpirun -hostfile myhostfile -np 14 ./a.out

will launch processes 0-3 on node aa, 4-7 on bb, and 8-11 on cc. It will then add the remaining two processes to whichever nodes it chooses.

One can also specify limits to oversubscription. For example, with the same hostfile:

mpirun -hostfile myhostfile -np 14 -nooversubscribe ./a.out

will produce an error since -nooversubscribe prevents oversubscription.

Limits to oversubscription can also be specified in the hostfile itself:
% cat myhostfile
aa slots=4 max_slots=4
bb max_slots=4
cc slots=4

The max_slots field specifies such a limit. When it does, the slots value defaults to the limit. Now:

mpirun -hostfile myhostfile -np 14 ./a.out

causes the first 12 processes to be launched as before, but the remaining two processes will be forced onto node cc. The other two nodes are protected by the hostfile against oversubscription by this job.

Using the --nooversubscribe option can be helpful since Open MPI currently does not get "max_slots" values from the resource manager.

Of course, -np can also be used with the -H or -host option. For example,

mpirun -H aa,bb -np 8 ./a.out

launches 8 processes. Since only two hosts are specified, after the first two processes are mapped, one to aa and one to bb, the remaining processes oversubscribe the specified hosts.

And here is a MIMD example:

mpirun -H aa -np 1 hostname : -H bb,cc -np 2 uptime

will launch process 0 running hostname on node aa and processes 1 and 2 each running uptime on nodes bb and cc, respectively.

Mapping, Ranking, and Binding: Oh My!

Open MPI employs a three-phase procedure for assigning process locations and ranks:

mapping

Assigns a default location to each process

ranking

Assigns an MPI_COMM_WORLD rank value to each process

binding

Constrains each process to run on specific processors

The mapping step is used to assign a default location to each process based on the mapper being employed. Mapping by slot, node, and sequentially results in the assignment of the processes to the node level. In contrast, mapping by object, allows the mapper to assign the process to an actual object on each node.

Note: the location assigned to the process is independent of where it will be bound - the assignment is used solely as input to the binding algorithm.

The mapping of process processes to nodes can be defined not just with general policies but also, if necessary, using arbitrary mappings that cannot be described by a simple policy. One can use the "sequential mapper," which reads the hostfile line by line, assigning processes to nodes in whatever order the hostfile specifies. Use the -mca rmaps seq option. For example, using the same hostfile as before:

mpirun -hostfile myhostfile -mca rmaps seq ./a.out

will launch three processes, one on each of nodes aa, bb, and cc, respectively. The slot counts don't matter; one process is launched per line on whatever node is listed on the line.

Another way to specify arbitrary mappings is with a rankfile, which gives you detailed control over process binding as well. Rankfiles are discussed below.

The second phase focuses on the ranking of the process within the job's MPI_COMM_WORLD. Open MPI separates this from the mapping procedure to allow more flexibility in the relative placement of MPI processes. This is best illustrated by considering the following two cases where we used the —map-by ppr:2:socket option:

node aa node bb

rank-by core 0 1 ! 2 3 4 5 ! 6 7

rank-by socket 0 2 ! 1 3 4 6 ! 5 7

rank-by socket:span 0 4 ! 1 5 2 6 ! 3 7

Ranking by core and by slot provide the identical result - a simple progression of MPI_COMM_WORLD ranks across each node. Ranking by socket does a round-robin ranking within each node until all processes have been assigned an MCW rank, and then progresses to the next node. Adding the span modifier to the ranking directive causes the ranking algorithm to treat the entire allocation as a single entity - thus, the MCW ranks are assigned across all sockets before circling back around to the beginning.

The binding phase actually binds each process to a given set of processors. This can improve performance if the operating system is placing processes suboptimally. For example, it might oversubscribe some multi-core processor sockets, leaving other sockets idle; this can lead processes to contend unnecessarily for common resources. Or, it might spread processes out too widely; this can be suboptimal if application performance is sensitive to interprocess communication costs. Binding can also keep the operating system from migrating processes excessively, regardless of how optimally those processes were placed to begin with.

The processors to be used for binding can be identified in terms of topological groupings - e.g., binding to an l3cache will bind each process to all processors within the scope of a single L3 cache within their assigned location. Thus, if a process is assigned by the mapper to a certain socket, then a —bind-to l3cache directive will cause the process to be bound to the processors that share a single L3 cache within that socket.

Alternatively, processes can be assigned to processors based on their local rank on a node using the --bind-to cpu-list:ordered option with an associated --cpu-list "0,2,5". In this example, the first process on a node will be bound to cpu 0, the second process on the node will be bound to cpu 2, and the third process on the node will be bound to cpu 5. --bind-to will also accept cpulist:ortered as a synonym to cpu-list:ordered. Note that an error will result if more processes are assigned to a node than cpus are provided.

To help balance loads, the binding directive uses a round-robin method when binding to levels lower than used in the mapper. For example, consider the case where a job is mapped to the socket level, and then bound to core. Each socket will have multiple cores, so if multiple processes are mapped to a given socket, the binding algorithm will assign each process located to a socket to a unique core in a round-robin manner.

Alternatively, processes mapped by l2cache and then bound to socket will simply be bound to all the processors in the socket where they are located. In this manner, users can exert detailed control over relative MCW rank location and binding.

Finally, --report-bindings can be used to report bindings.

As an example, consider a node with two processor sockets, each comprised of four cores, and each of those cores contains one hardware thread. We run mpirun with -np 4 --report-bindings and the following additional options:

% mpirun ... --map-by core --bind-to core
[...] ... binding child [...,0] to cpus 0001
[...] ... binding child [...,1] to cpus 0002
[...] ... binding child [...,2] to cpus 0004
[...] ... binding child [...,3] to cpus 0008

% mpirun ... --map-by socket --bind-to socket
[...] ... binding child [...,0] to socket 0 cpus 000f
[...] ... binding child [...,1] to socket 1 cpus 00f0
[...] ... binding child [...,2] to socket 0 cpus 000f
[...] ... binding child [...,3] to socket 1 cpus 00f0

% mpirun ... --map-by slot:PE=2 --bind-to core
[...] ... binding child [...,0] to cpus 0003
[...] ... binding child [...,1] to cpus 000c
[...] ... binding child [...,2] to cpus 0030
[...] ... binding child [...,3] to cpus 00c0

% mpirun ... --bind-to none

Here, --report-bindings shows the binding of each process as a mask. In the first case, the processes bind to successive cores as indicated by the masks 0001, 0002, 0004, and 0008. In the second case, processes bind to all cores on successive sockets as indicated by the masks 000f and 00f0. The processes cycle through the processor sockets in a round-robin fashion as many times as are needed.

In the third case, the masks show us that 2 cores have been bound per process. Specifically, the mapping by slot with the PE=2 qualifier indicated that each slot (i.e., process) should consume two processor elements. Since --use-hwthread-cpus was not specified, Open MPI defined "processor element" as "core", and therefore the --bind-to core caused each process to be bound to both of the cores to which it was mapped.

In the fourth case, binding is turned off and no bindings are reported.

Open MPI's support for process binding depends on the underlying operating system. Therefore, certain process binding options may not be available on every system.

Process binding can also be set with MCA parameters. Their usage is less convenient than that of mpirun options. On the other hand, MCA parameters can be set not only on the mpirun command line, but alternatively in a system or user mca-params.conf file or as environment variables, as described in the MCA section below. Some examples include:

mpirun option MCA parameter key value

--map-by core rmaps_base_mapping_policy core
--map-by socket rmaps_base_mapping_policy socket
--rank-by core rmaps_base_ranking_policy core
--bind-to core hwloc_base_binding_policy core
--bind-to socket hwloc_base_binding_policy socket
--bind-to none hwloc_base_binding_policy none

Rankfiles

Rankfiles are text files that specify detailed information about how individual processes should be mapped to nodes, and to which processor(s) they should be bound. Each line of a rankfile specifies the location of one process (for MPI jobs, the process' "rank" refers to its rank in MPI_COMM_WORLD). The general form of each line in the rankfile is:

rank <N>=<hostname> slot=<slot list>

For example:

$ cat myrankfile
rank 0=aa slot=1:0-2
rank 1=bb slot=0:0,1
rank 2=cc slot=1-2
$ mpirun -H aa,bb,cc,dd -rf myrankfile ./a.out

Means that

Rank 0 runs on node aa, bound to logical socket 1, cores 0-2.
Rank 1 runs on node bb, bound to logical socket 0, cores 0 and 1.
Rank 2 runs on node cc, bound to logical cores 1 and 2.

Rankfiles can alternatively be used to specify physical processor locations. In this case, the syntax is somewhat different. Sockets are no longer recognized, and the slot number given must be the number of the physical PU as most OS's do not assign a unique physical identifier to each core in the node. Thus, a proper physical rankfile looks something like the following:

$ cat myphysicalrankfile
rank 0=aa slot=1
rank 1=bb slot=8
rank 2=cc slot=6

This means that

Rank 0 will run on node aa, bound to the core that contains physical PU 1
Rank 1 will run on node bb, bound to the core that contains physical PU 8
Rank 2 will run on node cc, bound to the core that contains physical PU 6

Rankfiles are treated as logical by default, and the MCA parameter rmaps_rank_file_physical must be set to 1 to indicate that the rankfile is to be considered as physical.

The hostnames listed above are "absolute," meaning that actual resolveable hostnames are specified. However, hostnames can also be specified as "relative," meaning that they are specified in relation to an externally-specified list of hostnames (e.g., by mpirun's --host argument, a hostfile, or a job scheduler).

The "relative" specification is of the form "+n<X>", where X is an integer specifying the Xth hostname in the set of all available hostnames, indexed from 0. For example:

$ cat myrankfile
rank 0=+n0 slot=1:0-2
rank 1=+n1 slot=0:0,1
rank 2=+n2 slot=1-2
$ mpirun -H aa,bb,cc,dd -rf myrankfile ./a.out

Starting with Open MPI v1.7, all socket/core slot locations are be specified as logical indexes (the Open MPI v1.6 series used physical indexes). You can use tools such as HWLOC's "lstopo" to find the logical indexes of socket and cores.

Application Context or Executable Program?

To distinguish the two different forms, mpirun looks on the command line for --app option. If it is specified, then the file named on the command line is assumed to be an application context. If it is not specified, then the file is assumed to be an executable program.

Locating Files

If no relative or absolute path is specified for a file, Open MPI will first look for files by searching the directories specified by the --path option. If there is no --path option set or if the file is not found at the --path location, then Open MPI will search the user's PATH environment variable as defined on the source node(s).

If a relative directory is specified, it must be relative to the initial working directory determined by the specific starter used. For example when using the rsh or ssh starters, the initial directory is $HOME by default. Other starters may set the initial directory to the current working directory from the invocation of mpirun.

Current Working Directory

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

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

If the -wdir option is specified, Open MPI will attempt to change to the specified directory on all of the remote nodes. If this fails, mpirun will abort.

If the -wdir option is not specified, Open MPI will send the directory name where mpirun was invoked to each of the remote nodes. The remote nodes will try to change to that directory. If they are unable (e.g., if the directory does not exist on that node), then Open MPI will use the default directory determined by the starter.

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

Standard I/O

Open MPI directs UNIX standard input to /dev/null on all processes except the MPI_COMM_WORLD rank 0 process. The MPI_COMM_WORLD rank 0 process inherits standard input from mpirun. Note: The node that invoked mpirun need not be the same as the node where the MPI_COMM_WORLD rank 0 process resides. Open MPI handles the redirection of mpirun's standard input to the rank 0 process.

Open MPI directs UNIX standard output and error from remote nodes 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 Open MPI applications by using the typical shell redirection procedure on mpirun.

% mpirun -np 2 my_app < my_input > my_output

Note that in this example only the MPI_COMM_WORLD rank 0 process 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.

Signal Propagation

When orterun receives a SIGTERM and SIGINT, it will attempt to kill the entire job by sending all processes in the job a SIGTERM, waiting a small number of seconds, then sending all processes in the job a SIGKILL.

SIGUSR1 and SIGUSR2 signals received by orterun are propagated to all processes in the job.

A SIGTSTOP signal to mpirun will cause a SIGSTOP signal to be sent to all of the programs started by mpirun and likewise a SIGCONT signal to mpirun will cause a SIGCONT sent.

Other signals are not currently propagated by orterun.

Process Termination / Signal Handling

During the run of an MPI application, if any process 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.

User signal handlers should probably avoid trying to cleanup MPI state (Open MPI is currently not async-signal-safe; see MPI_Init_thread(3) for details about MPI_THREAD_MULTIPLE and thread safety). For example, if a segmentation 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 Open 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.

Process Environment

Processes in the MPI application inherit their environment from the Open RTE daemon upon the node on which they are running. The environment is typically inherited from the user's shell. On remote nodes, the exact environment is determined by the boot MCA module used. The rsh launch module, for example, uses either rsh/ssh to launch the Open RTE daemon on remote nodes, and typically executes one or more of the user's shell-setup files before launching the Open RTE 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 Open MPI.

See the "Remote Execution" section for more details.

Remote Execution

Open MPI requires that the PATH environment variable be set to find executables on remote nodes (this is typically only necessary in rsh- or ssh-based environments -- batch/scheduled environments typically copy the current environment to the execution of remote jobs, so if the current environment has PATH and/or LD_LIBRARY_PATH set properly, the remote nodes will also have it set properly). If Open MPI was compiled with shared library support, it may also be necessary to have the LD_LIBRARY_PATH environment variable set on remote nodes as well (especially to find the shared libraries required to run user MPI applications).

However, it is not always desirable or possible to edit shell startup files to set PATH and/or LD_LIBRARY_PATH. The --prefix option is provided for some simple configurations where this is not possible.

The --prefix option takes a single argument: the base directory on the remote node where Open MPI is installed. Open MPI will use this directory to set the remote PATH and LD_LIBRARY_PATH before executing any Open MPI or user applications. This allows running Open MPI jobs without having pre-configured the PATH and LD_LIBRARY_PATH on the remote nodes.

Open MPI adds the basename of the current node's "bindir" (the directory where Open MPI's executables are installed) to the prefix and uses that to set the PATH on the remote node. Similarly, Open MPI adds the basename of the current node's "libdir" (the directory where Open MPI's libraries are installed) to the prefix and uses that to set the LD_LIBRARY_PATH on the remote node. For example:

Local bindir:

/local/node/directory/bin

Local libdir:

/local/node/directory/lib64

If the following command line is used:

% mpirun --prefix /remote/node/directory

Open MPI will add "/remote/node/directory/bin" to the PATH and "/remote/node/directory/lib64" to the LD_LIBRARY_PATH on the remote node before attempting to execute anything.

The --prefix option is not sufficient if the installation paths on the remote node are different than the local node (e.g., if "/lib" is used on the local node, but "/lib64" is used on the remote node), or if the installation paths are something other than a subdirectory under a common prefix.

Note that executing mpirun via an absolute pathname is equivalent to specifying --prefix without the last subdirectory in the absolute pathname to mpirun. For example:

% /usr/local/bin/mpirun ...

is equivalent to

% mpirun --prefix /usr/local

Exported Environment Variables

All environment variables that are named in the form OMPI_* will automatically be exported to new processes on the local and remote nodes. Environmental parameters can also be set/forwarded to the new processes using the MCA parameter mca_base_env_list. The -x option to mpirun has been deprecated, but the syntax of the MCA param follows that prior example. While the syntax of the -x option and MCA param allows the definition of new variables, note that the parser for these options are currently not very sophisticated - it does not even understand quoted values. Users are advised to set variables in the environment and use the option to export them; not to define them.

Setting MCA Parameters

The -mca switch allows the passing of parameters to various MCA (Modular Component Architecture) modules. MCA modules have direct impact on MPI programs because they allow tunable parameters to be set at run time (such as which BTL communication device driver to use, what parameters to pass to that BTL, etc.).

The -mca switch takes two arguments: <key> and <value>. The <key> argument generally specifies which MCA module will receive the value. For example, the <key> "btl" is used to select which BTL to be used for transporting MPI messages. The <value> argument is the value that is passed. For example:

mpirun -mca btl tcp,self -np 1 foo

Tells Open MPI to use the "tcp" and "self" BTLs, and to run a single copy of "foo" an allocated node.

mpirun -mca btl self -np 1 foo

Tells Open MPI to use the "self" BTL, and to run a single copy of "foo" an allocated node.

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

Note that the -mca 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 Open MPI sets is:

OMPI_MCA_<key>=<value>

Thus, the -mca switch overrides any previously set environment variables. The -mca settings similarly override MCA parameters set in the $OPAL_PREFIX/etc/openmpi-mca-params.conf or $HOME/.openmpi/mca-params.conf file.

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 MCA module.

To find the available component types under the MCA architecture, or to find the available parameters for a specific component, use the ompi_info command. See the ompi_info(1) man page for detailed information on the command.

Setting MCA parameters and environment variables from file.

The -tune command line option and its synonym -mca mca_base_envar_file_prefix allows a user to set mca parameters and environment variables with the syntax described below. This option requires a single file or list of files separated by "," to follow.

A valid line in the file may contain zero or many "-x", "-mca", or “--mca” arguments. The following patterns are supported: -mca var val -mca var "val" -x var=val -x var. If any argument is duplicated in the file, the last value read will be used.

MCA parameters and environment specified on the command line have higher precedence than variables specified in the file.

Running as root

The Open MPI team strongly advises against executing mpirun as the root user. MPI applications should be run as regular (non-root) users.

Reflecting this advice, mpirun will refuse to run as root by default. To override this default, you can add the --allow-run-as-root option to the mpirun command line, or you can set the environmental parameters OMPI_ALLOW_RUN_AS_ROOT=1 and OMPI_ALLOW_RUN_AS_ROOT_CONFIRM=1. Note that it takes setting two environment variables to effect the same behavior as --allow-run-as-root in order to stress the Open MPI team's strong advice against running as the root user. After extended discussions with communities who use containers (where running as the root user is the default), there was a persistent desire to be able to enable root execution of mpirun via an environmental control (vs. the existing --allow-run-as-root command line parameter). The compromise of using two environment variables was reached: it allows root execution via an environmental control, but it conveys the Open MPI team's strong recomendation against this behavior.

Exit status

There is no standard definition for what mpirun should return as an exit status. After considerable discussion, we settled on the following method for assigning the mpirun exit status (note: in the following description, the "primary" job is the initial application started by mpirun - all jobs that are spawned by that job are designated "secondary" jobs):

•

if all processes in the primary job normally terminate with exit status 0, we return 0

•

if one or more processes in the primary job normally terminate with non-zero exit status, we return the exit status of the process with the lowest MPI_COMM_WORLD rank to have a non-zero status

•

if all processes in the primary job normally terminate with exit status 0, and one or more processes in a secondary job normally terminate with non-zero exit status, we (a) return the exit status of the process with the lowest MPI_COMM_WORLD rank in the lowest jobid to have a non-zero status, and (b) output a message summarizing the exit status of the primary and all secondary jobs.

•

if the cmd line option --report-child-jobs-separately is set, we will return -only- the exit status of the primary job. Any non-zero exit status in secondary jobs will be reported solely in a summary print statement.

By default, the job will abort when any process terminates with non-zero status. The MCA parameter "orte_abort_on_non_zero_status" can be set to "false" (or "0") to cause OMPI to not abort a job if one or more processes return a non-zero status. In that situation the OMPI records and notes that processes exited with non-zero termination status to report the approprate exit status of mpirun (per bullet points above).