This document describes how to run Chapel across multiple nodes of an InfiniBand cluster, including Cray CS and HPE Apollo systems. Multilocale Chapel Execution describes general information about running Chapel in a multilocale configuration.
Due to the wide variety of InfiniBand clusters 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 a Cray CS or HPE Apollo system
CHPL_HOST_PLATFORM can instead be set, in which case the comm
and substrate settings will be inferred.
For Cray CS:
For HPE Apollo:
gasnetrun_ibv-based launcher should be used to launch jobs and
native launchers like
srun will not work. Most InfiniBand
clusters 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:
run jobs interactively on your system
queue jobs using Slurm (srun/sbatch)
queue jobs using PBS (qsub)
queue jobs using LSF (bsub)
CHPL_LAUNCHER will typically default to
is in your path, in which case it will default to
By default Slurm, PBS, and LSF versions launch in an interactive
mode. For batch submission with Slurm
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.
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
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'.
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 to a small value can limit
communication performance so it is highly recommend to use the value
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 two primary means of doing this on InfiniBand clusters are
mpi. GASNet will default to
mpi if MPI support
is detected at configure time, otherwise it will default to
mpi will likely incur a performance penalty because MPI
will be running concurrently with GASNet. Running with
recommended, but not all systems support ssh’ing to compute nodes so
it is not always the default.
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.
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)