Using Chapel with InfiniBand ¶

This document describes how to run Chapel across multiple nodes of an InfiniBand system, including HPE Apollo and Cray CS systems. Multilocale Chapel Execution describes general information about running Chapel in a multilocale configuration.

Configuring for InfiniBand ¶

Due to the wide variety of InfiniBand systems and potential for false-positives, Chapel does not currently auto-detect InfiniBand configurations or platforms that commonly use InfiniBand. To build Chapel with InfiniBand support, set:

export CHPL_COMM=gasnet
export CHPL_COMM_SUBSTRATE=ibv

Alternatively, when running on an HPE Apollo or Cray CS system CHPL_HOST_PLATFORM can be set instead, in which case the comm and substrate settings will be inferred.

For HPE Apollo:

export CHPL_HOST_PLATFORM=hpe-apollo

For Cray CS:

export CHPL_HOST_PLATFORM=cray-cs

UCX alternative ¶

In rare cases, we have had users who are unable to use the ibv conduit on their InfiniBand systems. In such cases, the ucx conduit is an alternative to consider, which requires setting:

export CHPL_COMM=gasnet
export CHPL_COMM_SUBSTRATE=ucx

(and, most likely, export CHPL_LAUNCHER=gasnetrun_ucx when you reach the following section).

Note that we have not had as much experience with the ucx conduit in Chapel’s development and testing as we have with ibv, so it is currently considered something of an experimental feature. In addition, the GASNet team also currently classifies the ucx conduit as “experimental” in the sense that it is feature complete but has not received as much performance-oriented focus as ibv has.

Configuring a Launcher ¶

A gasnetrun_ibv-based launcher should be used to launch jobs and native launchers like srun will not work. Most InfiniBand systems use a workload manager or queueing system such as Slurm, LSF, or PBS. To select an appropriate Chapel launcher you can set CHPL_LAUNCHER to one of the following values:

Launcher Name	Description
gasnetrun_ibv	run jobs interactively on your system
slurm-gasnetrun_ibv	queue jobs using Slurm (srun/sbatch)
pbs-gasnetrun_ibv	queue jobs using PBS (qsub)
lsf-gasnetrun_ibv	queue jobs using LSF (bsub)

CHPL_LAUNCHER will typically default to gasnetrun_ibv unless CHPL_HOST_PLATFORM is cray-cs or hpe-apollo and srun is in your path, in which case it will default to slurm-gasnetrun_ibv

By default Slurm, PBS, and LSF versions launch in an interactive mode. For batch submission with Slurm CHPL_LAUNCHER_USE_SBATCH can be used as described in Using Slurm. For other launchers and as an alternative for Slurm, users can write batch submission scripts and use gasnetrun_ibv to launch their jobs.

Selecting a Spawner ¶

Under the covers, gasnetrun_ibv-based launchers must figure out how to spawn jobs and get them up and running on the compute nodes. GASNet’s three ways of doing this on InfiniBand systems are ssh, pmi, and mpi, described just below. When MPI is detected at configure time, it will be selected as the default, but we recommend using one of the other options if possible. This can be done by setting the GASNET_IBV_SPAWNER environment variable, whose legal values are:

ssh: Based on our experience, this is the preferred option, when possible. This requires the ability to ssh to the system’s compute nodes, which is not supported by all systems, depending on how they are configured. See the following sub-section for details on this option.
pmi: When GASNet’s configure step detects a PMI-capable job scheduler like Slurm, pmi can be the next best choice because it often “just works” and can reduce overhead compared to mpi. For more information about this option, including how to configure job launch via PMIRUN_CMD, see the GASNet README for the PMI-based spawner (also available at $CHPL_HOME/third-party/gasnet/gasnet-src/other/pmi-spawner/README).
mpi: When the previous cases are not options, mpi serves as a reasonable last resort. Note that it may, depending on its configuration, incur a performance penalty due to competition between MPI and GASNet for limited communication resources. See the Using MPI for Job Launch section below for best practices when using this option.

Using SSH for Job Launch ¶

To launch InfiniBand jobs with SSH, use the following:

# Specify ssh spawner
export GASNET_IBV_SPAWNER=ssh

# Specify the nodes to run on (only required when using plain
# gasnetrun_ibv outside a Slurm/PBS/LSF reservation)
export GASNET_SSH_SERVERS="nid00001 nid00002 nid00003 ..."

If you receive an error message like:

*** Failed to start processes on nid00001, possibly due to an inability to establish an ssh connection from login-node without interactive authentication.

This indicates passwordless SSH is not set up. You can try copying existing SSH keys or generating new ones with the following:

ssh-keygen -t rsa # use default location and empty passphrase
cat ~/.ssh/id_rsa.pub >> ~/.ssh/authorized_keys

If you see the same error message this may indicate ssh connections to compute nodes are not allowed, in which case using the MPI spawner may be your only option.

For further information about environment variables that can be used to control how ssh is used to launch your Chapel program, see the descriptions of GASNET_SSH_CMD and GASNET_SSH_OPTIONS in the GASNet README for the ssh spawner (also available at $CHPL_HOME/third-party/gasnet/gasnet-src/other/ssh-spawner/README).

Using MPI for Job Launch ¶

To launch InfiniBand jobs with mpirun, first make sure that mpicc is available and that MPI programs launch appropriately with mpirun. Then use the following. You’ll want to make sure that GASNet detects MPI in its configuration output.

export GASNET_IBV_SPAWNER=mpi

As mentioned above, a potential downside of using MPI for launching Chapel programs is that the resources that it requires to get the program up and running can interfere with those needed by GASNet. In some cases, this can result in negative impacts on performance. In others, it can prevent GASNet from accessing the network resources it requires at all. For example, the following error is an example of one in which MPI is preventing GASNet from accessing what it needs (albeit using the ofi conduit rather than the ibv conduit preferred for InfiniBand):

*** FATAL ERROR (proc 0): in gasnetc_ofi_init() at /third-party/gasnet/gasnet-src/ofi-conduit/gasnet_ofi.c:1336: fi_endpoint for rdma failed: -22(Invalid argument)

For tips and best practices about how to configure/use GASNet to avoid such conflicts with MPI, please refer to the section “Build-time Configuration” in the GASNet README for the MPI spawner (also available at $CHPL_HOME/third-party/gasnet/gasnet-src/other/mpi-spawner/README). Within this README, see also the description of the MPIRUN_CMD environment variable as a means of configuring how jobs are started.

Verifying Job Launch ¶

Once the above configuration has been done, checking that jobs are launching properly is recommended. The following Chapel program will print out the locale names and how much parallelism is available per locale. Ideally each locale is running on a unique node (not oversubscribed) and the amount of parallelism matches the number of physical cores on each node.

for loc in Locales do on loc do
  writeln((here.name, here.maxTaskPar));

An example run may look something like the following:

(nid00001, 28)
(nid00002, 28)

If nodes are oversubscribed or the amount of parallelism is far less than expected see Selecting a Spawner and if that does not help consider opening a bug as described in Reporting Chapel Issues.

Setting Memory Registration Limits ¶

On most high-performance networks, including InfiniBand, memory has to be registered with the network in order for Chapel to take advantage of fast one-sided communication. On InfiniBand networks there may be limits placed on how much memory can be registered so GASNet will probe at startup to detect this value. This probing can be slow, so GASNet will recommend setting GASNET_PHYSMEM_MAX to avoid probing every time. On nodes with homogeneous amounts of memory this message usually looks something like:

WARNING: Beginning a potentially slow probe of max pinnable memory...
WARNING: Probe of max pinnable memory completed in 45s.
WARNING:   Probe of max pinnable memory has yielded '335 GB'.
WARNING:   If you have the same memory configuration on all nodes, then
WARNING:   to avoid this probe in the future either reconfigure using
WARNING:      --with-ibv-physmem-max='335 GB'
WARNING:   or run with environment variable
WARNING:      GASNET_PHYSMEM_MAX='335 GB'.

Where setting GASNET_PHYSMEM_MAX='335 GB' will quiet the warning and skip the startup probe. On nodes with non-homogeneous amounts of memory GASNet may recommend using a fraction of memory instead of an absolute value with something like GASNET_PHYSMEM_MAX='0.667'.

Setting GASNET_PHYSMEM_MAX to a small value can limit communication performance so it is highly recommended to use the value GASNet suggests.

Co-locales ¶

By default, Chapel disables GASNet’s PSHM (Process SHared-Memory bypass) feature when running on InfiniBand. This means that by default on-node communication between co-locales traverses the loopback IB network interface, incurring overheads associated with the IB verbs networking layer that are high relative to the latencies one might expect for on-node communication. Using co-locales in this configuration will generate a startup warning from GASNet that looks like this:

*** WARNING (proc 0): Running with multiple processes per host without shared-memory communication support (PSHM).  This can significantly reduce performance.  Please re-configure GASNet using `--enable-pshm` to enable fast intra-host comms.

This (somewhat confusingly worded) message accurately reflects the fact that Chapel’s co-locale behavior with the InfiniBand backend has not yet been tuned and may provide sub-optimal performance for on-node communication.

Users wishing to experiment with enabling the shared-memory bypass support for the InfiniBand backend can do so by adding the following environment variable setting when building Chapel:

export CHPL_GASNET_MORE_CFG_OPTIONS=--enable-pshm

This will enable the PSHM support when co-locales are in-use, such that on-node communication between co-locales will occur in memory and not through the network. It’s worth noting that Chapel’s integration of the PSHM feature currently requires an extra “progress” thread that defaults to running on the same cores as the Chapel tasks, and will compete with those tasks for cycles. For computation-bound applications the overhead incurred by this progress thread may outweigh the benefits of PSHM. You can optionally disable PSHM at application run-time by setting export GASNET_SUPERNODE_MAXSIZE=1, although this won’t exactly match the behavior of building without PSHM support.

Another alternative is to dedicate a core for the progress thread, preventing it from running on the same cores as the Chapel tasks. This is accomplished by setting CHPL_RT_COMM_GASNET_DEDICATED_PROGRESS_CORE=true. Note that this means there will be one fewer core to run Chapel tasks, which may not be advantageous on machines with relatively few cores. Also note that this variable will dedicate a core whether or not PSHM is in-use, so you should probably only set this variable if you are using co-locales and PSHM is enabled.

Progress Threads ¶

In addition to the “busy” polling-based progress thread described above that is enabled when co-locales are combined with GASNet’s (default disabled) PSHM support, the InfiniBand backend also optionally includes GASNet-level blocking progress threads used to help retire incoming and outgoing network communication operations (named the blocking “receive” and “send” progress threads, respectively). These threads are “blocking” in that they are awakened on-demand when the network adapter reports there is communication work to be done, and otherwise generally remain parked in a kernel call where they do not consume any core cycles. By default the blocking receive progress thread is enabled and the blocking send progress thread is disabled.

The blocking send progress thread may be enabled at application run-time by setting:

export GASNET_SND_THREAD=1

This enables a helper thread that has been shown to accelerate injection of communication in some cases, but in other cases may degrade communication throughput (notably on NUMA systems where the locale straddles the NUMA boundary). GASNet also provides additional environment variable settings that can optionally be used to control the detailed behavior of these threads, see the GASNet documentation referenced in See Also.

By default, the blocking progress threads are created by GASNet and do not have any thread-specific core binding. An experimental Chapel feature allows more control over the behavior of the blocking progress threads:

export CHPL_RT_COMM_GASNET_DEFER_PROGRESS_THREADS=true

Specifically, this setting enables the CHPL_RT_COMM_GASNET_DEDICATED_PROGRESS_CORE setting described in the earlier section to additionally control the placement and binding of GASNet’s blocking send and receive progress threads.