Chapel Launchers

When compiling Chapel programs for multiple locales, a launcher binary is typically created that will execute the appropriate command(s) to get your program started. For example, when compiling for multiple locales, typically two binaries will be generated by the compiler (e.g., myprogram and myprogram_real). The first binary contains code to get your program up and running on multiple locales while the second contains your actual program code.

The goals of the launcher binary are:

  1. to wrap details of job startup in a portable way so that new users can quickly get Chapel programs up and running on an unfamiliar platform.

  2. to perform command-line parsing and error checking prior to waiting in a queue or firing off a parallel job in order to save time and resources related to simple errors/typos in the command line.

  3. to preserve Chapel’s global-view programming model by permitting the user to run their program using a single binary (corresponding to the single logical task that executes main()) without getting bogged down in questions of numbers of nodes, numbers of cores per node, numbers of program instances to start up, etc.

  4. if necessary, to coordinate runtime functional activity, such as I/O.

Executing a Chapel program using the verbose (-v) flag will typically print out the command(s) used to launch the program, along with any environment variables the launcher set on its behalf. It will also cause the program itself to print additional information about how it configured itself, though most of this will be of more interest to Chapel developers than regular users.

Executing using the help (-h/--help) flag will typically print out any launcher-specific options in addition to the normal help message for the program itself.

You can also execute the Chapel launcher with the --dry-run flag. This will not actually run or launch the user program, but instead simply print the same thing as -v: the command(s) that would have been used to launch the program, along with any environment variables the launcher would have set on its behalf. Note that --dry-run will also cause batch and other files created for the system launcher to be left behind so you can inspect and/or reuse them. Normally these are removed by the Chapel launcher when the program finishes. An example of such a file would be the sbatch file created when a Slurm-based launcher is used and CHPL_LAUNCHER_USE_SBATCH is set.

Currently Supported Launchers

Currently supported launchers include:

Launcher Name

Description

amudprun

GASNet launcher for the UDP substrate

aprun

Cray application launcher using aprun

gasnetrun_ibv

GASNet launcher for the Infiniband substrate

gasnetrun_mpi

GASNet launcher for the MPI substrate

mpirun4ofi

provisional launcher for CHPL_COMM=ofi on non-Cray systems

lsf-gasnetrun_ibv

GASNet launcher for LSF (bsub) and the Infiniband substrate

pals

Cray application launcher for PALS on HPE Cray EX systems

pbs-aprun

Cray application launcher for PBS (qsub) + aprun

pbs-gasnetrun_ibv

GASNet launcher for PBS (qsub) and the Infiniband substrate

slurm‑gasnetrun_ibv

GASNet launcher for SLURM and the Infiniband substrate

slurm‑gasnetrun_mpi

GASNet launcher for SLURM and the MPI substrate

slurm‑gasnetrun_ofi

GASNet launcher for SLURM and the OFI substrate

slurm-srun

native SLURM launcher

smp

GASNet launcher for the shared-memory substrate

none

do not use a launcher

A specific launcher can be explicitly requested by setting the CHPL_LAUNCHER environment variable. For the specific case of the mpirun4ofi launcher, please see Using Chapel with libfabric.

If CHPL_LAUNCHER is left unset, a default is picked as follows:

  • if CHPL_COMM is gasnet and CHPL_COMM_SUBSTRATE is udp CHPL_LAUNCHER is set to amudprun

  • otherwise, if CHPL_TARGET_PLATFORM is cray-xc or hpe-cray-ex:

    If

    CHPL_LAUNCHER

    both aprun and srun in user’s path

    none

    aprun in user’s path

    aprun

    srun in user’s path

    slurm-srun

    otherwise

    none

  • otherwise, if CHPL_TARGET_PLATFORM is cray-cs and CHPL_COMM is gasnet and salloc is in the user’s path:

    If

    CHPL_LAUNCHER

    CHPL_COMM_SUBSTRATE=ibv

    slurm-gasnetrun_ibv

    CHPL_COMM_SUBSTRATE=mpi

    slurm-gasnetrun_mpi

  • otherwise, if CHPL_TARGET_PLATFORM is cray-cs and srun is in the users path CHPL_LAUNCHER is set to slurm-srun

  • otherwise, if CHPL_COMM is gasnet:

    If

    CHPL_LAUNCHER

    CHPL_COMM_SUBSTRATE=ibv

    gasnetrun_ibv

    CHPL_COMM_SUBSTRATE=mpi

    gasnetrun_mpi

    CHPL_COMM_SUBSTRATE=smp

    smp

    otherwise

    none

  • otherwise CHPL_LAUNCHER is set to none

If the launcher binary does not work for your system (due to an installation-specific configuration, e.g.), you can often use the --dry-run flag to capture the commands that the launcher would have executed on your behalf and customize them for your needs.

Forwarding Environment Variables

Chapel launchers generally arrange for environment variables to be forwarded to worker processes. However, this strategy is not always reliable. The remote system may override some environment variables, and some launchers might not correctly forward all environment variables.

CHPL_RT_MASTERIP

This environment variable is used to specify the IP address which should be used to connect. By default, the node creating the connection will pass the result of gethostname() on to the nodes that need to connect to it, which will resolve that to an IP address using gethostbyname().

When CHPL_COMM == gasnet, this will also be used to set the value of GASNET_MASTERIP, which corresponds to the hostname of the master node (see https://gasnet.lbl.gov/dist/udp-conduit/README ).

CHPL_RT_WORKERIP

This environment variable is used to specify the IP address which should be used to communicate between worker nodes. By default, worker nodes will communicate among themselves using the same interface used to connect to the master node (see CHPL_RT_MASTERIP, above).

When CHPL_COMM == gasnet, this will also be used to set the value of GASNET_WORKERIP (see https://gasnet.lbl.gov/dist/udp-conduit/README ).

Using Slurm

To use native Slurm, set:

export CHPL_LAUNCHER=slurm-srun

On Cray systems, this will happen automatically if srun is found in your path, but not when both srun and aprun are found in your path. Native Slurm is the best option where it works, but at the time of this writing, there are problems with it when combined with CHPL_COMM=gasnet and the UDP or InfiniBand conduits. So, for these configurations please see:

Common Slurm Settings

  • Optionally, you can specify a node access mode by setting the environment variable CHPL_LAUNCHER_NODE_ACCESS. It will default to exclusive access, but can be overridden to:

    • shared to give shared access to nodes

    • unset to use the system default and not specify a node access mode

    • exclusive to give exclusive access to nodes (this is the default)

    For example, to grant shared node access, set:

    export CHPL_LAUNCHER_NODE_ACCESS=shared

  • Optionally, you can specify a slurm partition by setting the environment variable CHPL_LAUNCHER_PARTITION. For example, to use the ‘debug’ partition, set:

    export CHPL_LAUNCHER_PARTITION=debug

  • Optionally, you can specify a slurm nodelist by setting the environment variable CHPL_LAUNCHER_NODELIST. For example, to use node nid00001, set:

    export CHPL_LAUNCHER_NODELIST=nid00001

  • Optionally, you can specify a slurm constraint by setting the environment variable CHPL_LAUNCHER_CONSTRAINT. For example, to use nodes with the ‘cal’ feature (as defined in the slurm.conf file), set:

    export CHPL_LAUNCHER_CONSTRAINT=cal

  • Optionally, you can specify a slurm account by setting the environment variable CHPL_LAUNCHER_ACCOUNT. For example, to use the account ‘acct’, set:

    export CHPL_LAUNCHER_ACCOUNT=acct

  • If the environment variable CHPL_LAUNCHER_USE_SBATCH is defined then sbatch is used to launch the job to the queue system, rather than running it interactively as usual. In this mode, the output will be written by default to a file called <executableName>.<jobID>.out. The environment variable CHPL_LAUNCHER_SLURM_OUTPUT_FILENAME can be used to specify a different filename for the output.

Using any SSH-based launcher with Slurm

It is possible to use any SSH-based launcher with Slurm, with some additionally effort. This strategy can come in handy if other launchers are not working. However, launchers such as slurm-srun and slurm-gasnetrun_ibv offer a better experience.

First, let’s see how to use an SSH-based launcher with an interactive salloc session. Here we will assume the UDP conduit, but any other launcher supporting SSH can be configured analogously.

# Compile a sample program
chpl -o hello6-taskpar-dist examples/hello6-taskpar-dist.chpl

# Reserve 2 nodes for an interactive run
salloc -N 2
# Then, within the salloc shell

  # Specify that ssh should be used
  export GASNET_SPAWNFN=S
  # Specify the list of nodes to use; SSH_SERVERS can also be used
  export GASNET_SSH_SERVERS=`scontrol show hostnames | xargs echo`
  # Run the program on the 2 reserved nodes.
  ./hello6-taskpar-dist -nl 2

This strategy can also be used within an sbatch script. Here is an example script to save to the file job.bash:

#!/bin/bash
#SBATCH -t 0:10:0
#SBATCH --nodes=2
#SBATCH --exclusive
#SBATCH --partition=chapel
#SBATCH --output=job.output

export GASNET_SPAWNFN=S
export GASNET_SSH_SERVERS=`scontrol show hostnames | xargs echo`

./hello6-taskpar-dist -nl 2

To run this job, use:

sbatch job.bash

and when it completes, the output will be available in job.output as specified in job.bash.

Changing the _real binary suffix

In order to support profiling tools that produce new binaries for the launcher to execute, the suffix of the real binary executed by the launcher may be changed with the CHPL_LAUNCHER_SUFFIX environment variable. If this variable is unset, the suffix defaults to “_real”, matching the compiler’s output.

Bypassing the launcher

If the Chapel launcher capability fails you completely, set CHPL_LAUNCHER to none, recompile, and execute the resulting binary according to the following rules using tools and queueing mechanisms appropriate for your system:

  • on most systems, the number of locales should be equal to the number of nodes on which you execute. That in turn should match the number of copies of the program that you are running.

  • some queueing systems require you to specify the number of cores to use per node. For best results, you will typically want to use all of them. All intra-node parallelism is typically implemented using Chapel’s threading layer (e.g., pthreads), so extra copies of the binary are not required per core.

  • in our experience, this technique does not work for InfiniBand configurations.

Additional launchers

In addition to the supported launchers listed above there are several others that are not actively maintained but may still work.

Launcher Name

Description

mpirun

launch using mpirun (no mpi comm currently)