Using Chapel on HPE Cray Systems 

The following information is assembled to help Chapel users get up and running on HPE Cray systems including the HPE Cray EX and Cray XC series systems.

Getting Started with Chapel on HPE Cray EX or XC Systems 

Chapel is available as a module for HPE Cray EX and XC systems. When it is installed on your system, you do not need to build Chapel from the source release (though you can). Using the module systems on such platforms, you can use Chapel with the default settings and confirm it is correctly installed, as follows:

Ensure this required module is loaded. Normally it will be loaded for you, but under some circumstances you may need to load or restore it yourself:
```
PrgEnv-gnu or PrgEnv-cray
```
Load the Chapel module:
```
module load chapel
```

Compile an example program using:

chpl $CHPL_HOME/examples/hello6-taskpar-dist.chpl

Execute the resulting executable (on four locales):
```
./hello6-taskpar-dist -nl 4
```

This should be all that is necessary to use Chapel on an HPE Cray EX or XC system. If the installation setup by your system administrator deviates from the default settings, or you are interested in other configuration options, see Using Chapel on an HPE Cray System below. If instead you wish to build Chapel from source, continue on to Building Chapel for an HPE Cray System from Source just below.

Note that the Chapel module for HPE Cray EX systems does not currently have the GASNet communication layer built into it. For multilocale execution on EX systems please use the ofi communication layer instead.

For information on obtaining and installing the Chapel module please contact your system administrator.

Getting Started with Chapel on HPE Apollo and Cray CS Systems 

On HPE Apollo and Cray CS systems, Chapel is not currently available as a module due to the wide diversity of configurations that these systems support. For this reason, Chapel must be built from source on these systems using the Building Chapel for an HPE Cray System from Source instructions just below.

Building Chapel for an HPE Cray System from Source 

Set CHPL_HOST_PLATFORM to the string representing your system type.

For example:
export CHPL_HOST_PLATFORM=hpe-cray-ex
The following table lists the supported systems and strings. Note that on HPE Cray EX and XC systems, these values should typically be inferred automatically and not need to be set manually. That said, there is also no downside to setting them manually. As with most CHPL_* environment variables, the current set and inferred values can be determined by running $CHPL_HOME/util/printchplenv.

System

CHPL_HOST_PLATFORM

EX series

hpe-cray-ex

Apollo series

hpe-apollo

XD series

hpe-cray-xd

XC series

cray-xc

CS series

cray-cs

Optionally, set the CHPL_LAUNCHER environment variable to indicate how Chapel should launch jobs on your system:

On an HPE Cray EX system, …

set CHPL_LAUNCHER to…

…queue jobs using SLURM (sbatch)

slurm-srun

…run jobs interactively using PALS

pals

On a HPE Apollo or Cray CS system, to…

set CHPL_LAUNCHER to…

…run jobs interactively on your system

gasnetrun_ibv

…queue jobs using SLURM (sbatch)

slurm-gasnetrun_ibv

…queue jobs using LSF (bsub)

lsf-gasnetrun_ibv

…queue jobs using PBS (qsub)

pbs-gasnetrun_ibv

On a Cray XC system, to…

set CHPL_LAUNCHER to…

…run jobs interactively on your system

aprun

…queue jobs using PBS (qsub)

pbs-aprun

…queue jobs using SLURM (sbatch)

slurm-srun

On HPE Cray EX systems, CHPL_LAUNCHER defaults to slurm-srun if srun is in your path and none otherwise.

On HPE Apollo and Cray CS systems, CHPL_LAUNCHER defaults to gasnetrun_ibv.

On Cray XC systems, CHPL_LAUNCHER defaults to aprun if aprun is in your path, slurm-srun if srun is in your path and none otherwise.

You can also set CHPL_LAUNCHER to none if you prefer to manually manage all queuing and job launch commands yourself, though the advantage of using a launcher is to make sure that Chapel maps processes and threads to the appropriate target hardware units.

For more information on Chapel’s launcher capabilities and options, refer to Chapel Launchers.

Select the target compiler that Chapel should use when compiling code for the compute node:

On an HPE Apollo or Cray CS series system, set the CHPL_TARGET_COMPILER environment variable to indicate which compiler to use (and make sure that the compiler is in your path).

To request…

set CHPL_TARGET_COMPILER to…

…the LLVM/clang backend

llvm (default)

…the GNU compiler (gcc)

gnu

…the Clang compiler

clang

…the Intel compiler (icc)

intel

On an HPE Cray EX or Cray XC system, when using the C back-end, ensure that you have one of the following Programming Environment modules loaded to specify your target compiler:
```
PrgEnv-allinea (ARM only)
PrgEnv-cray
PrgEnv-gnu
PrgEnv-intel
```
Make sure you’re in the top-level chapel/ directory and make/re-make the compiler and runtime:
```
gmake
```
Note that a single Chapel installation can support multiple configurations simultaneously and that you can switch between them simply by changing any of the above settings. However, each configuration must be built separately. Thus, you can change any of the settings in the steps before this, and then re-run this step in order to create additional installations. Thereafter, you can switch between any of these configurations without rebuilding.

Using Chapel on an HPE Cray System 

If you are working from a Chapel module:
1. Load the module using module load chapel
2. Optionally select a launcher, as in step 2 above
3. Select a target compiler, as in step 3 above
If you are working from a source installation:
1. Set your host platform as in step 1 above
2. Optionally select a launcher, as in step 2 above
3. Select a target compiler, as in step 3 above
4. Set CHPL_HOME and your paths by invoking the appropriate util/setchplenv script for your shell. For example:
source util/setchplenv.bash
On HPE Cray EX systems with CHPL_COMM=ofi, optionally, load the Cray PMI modules:
```
module load cray-pmi{,-lib}
```
Often this is not required. Usually the default PMI support has sufficient capabilities to support Chapel program startup. But under certain circumstances it does not, and when you run a Chapel program that was built without these loaded you will see messages like this one:
```
[PE_0]:_pmi2_add_kvs:ERROR: The KVS data segment of <num> entries
is not large enough.  Increase the number of KVS entries by
setting env variable PMI_MAX_KVS_ENTRIES to a higher value.
```
Having the Cray PMI modules loaded when the program is compiled will prevent this problem. We expect that eventually these modules will be loaded by default on EX systems, but so far this has not consistently been the case.
Compile your Chapel program. For example:
```
chpl $CHPL_HOME/examples/hello6-taskpar-dist.chpl
```
See Compiling Chapel Programs or man chpl for further details.
If CHPL_LAUNCHER is set to anything other than none, when you compile a Chapel program for your Cray system, you will see two binaries (e.g., hello6-taskpar-dist and hello6-taskpar-dist_real). The first binary contains code to launch the Chapel program onto the compute nodes, as specified by your CHPL_LAUNCHER setting. The second contains the program code itself; it is not intended to be executed directly from the shell prompt.

You can use the -v or --dry-run flags to see the commands used by the launcher binary to start your program.

If CHPL_LAUNCHER is pbs-aprun:
1. You can optionally specify a queue name using the environment variable CHPL_LAUNCHER_QUEUE. For example:
  export CHPL_LAUNCHER_QUEUE=batch
  If this variable is left unset, no queue name will be specified. Alternatively, you can set the queue name on your Chapel program command line using the --queue flag.
2. You can also optionally set a wall clock time limit for the job using CHPL_LAUNCHER_WALLTIME. For example to specify a 10-minute time limit, use:
  export CHPL_LAUNCHER_WALLTIME=00:10:00
  Alternatively, you can set the wall clock time limit on your Chapel program command line using the --walltime flag.
For further information about launchers, please refer to Chapel Launchers.
Execute your Chapel program. Multi-locale executions require the number of locales (compute nodes) to be specified on the command line. For example:
```
./hello6-taskpar-dist -nl 2
```
Requests the program to be executed using two locales.
If your HPE Cray system has compute nodes with varying numbers of cores, you can request nodes with at least a certain number of cores using the variable CHPL_LAUNCHER_CORES_PER_LOCALE. For example, on a system in which some compute nodes have 24 or more cores per compute node, you could request nodes with at least 24 cores using:
```
export CHPL_LAUNCHER_CORES_PER_LOCALE=24
```
This variable may be needed when you are using the aprun launcher and running Chapel programs within batch jobs you are managing yourself. The aprun launcher currently creates aprun commands that request the maximum number of cores per locale found on any locale in the system, irrespective of the fact that the batch job may have a lower limit than that on the number of cores per locale. If the batch job limit is less than the maximum number of cores per locale, you will get the following error message when you try to run a Chapel program:
```
apsched: claim exceeds reservation's CPUs
```
You can work around this by setting CHPL_LAUNCHER_CORES_PER_LOCALE to the same or lesser value as the number of cores per locale specified for the batch job (for example, the mppdepth resource for the PBS qsub command). In the future we hope to achieve better integration between Chapel launchers and workload managers.
If your HPE Cray system has compute nodes with varying numbers of CPUs per compute unit, you can request nodes with a certain number of CPUs per compute unit using the variable CHPL_LAUNCHER_CPUS_PER_CU. For example, on a Cray XC system with some nodes having at least 2 CPUs per compute unit, to request running on those nodes you would use:
```
export CHPL_LAUNCHER_CPUS_PER_CU=2
```
Currently, the only legal values for CHPL_LAUNCHER_CPUS_PER_CU are 0 (the default), 1, and 2.

For more information on…	see…
…CHPL_* environment settings	Setting up Your Environment for Chapel
…Compiling Chapel programs	Compiling Chapel Programs
…Launcher options	Chapel Launchers
…Executing Chapel programs	Executing Chapel Programs
…Running multi-locale Chapel programs	Multilocale Chapel Execution

Cray File Systems and Chapel execution 

For best results, it is recommended that you execute your Chapel program by placing the binaries on a file system shared between the login node and compute nodes (typically Lustre), as this will provide the greatest degree of transparency when executing your program. In some cases, running a Chapel program from a non-shared file system will make it impossible to launch onto the compute nodes. In other cases, the launch will succeed, but any files read or written by the Chapel program will be opened relative to the compute node’s file system rather than the login node’s.

Special Notes for Cray XC Series Systems 

Native ugni Communication Layer 

The native ugni communication layer is the default on Cray XC systems. See Using Chapel with ugni for more information on using the ugni communication layer.

GASNet Communication Layer 

The GASNet-based communication layer discussed in the Using Chapel with GASNet page can be used on Cray XC systems, using the ‘ofi’ or ‘mpi’ substrates. This is not well supported or tested. The relevant configurations are:

CHPL_COMM=gasnet
CHPL_COMM_SUBSTRATE=ofi or mpi CHPL_GASNET_SEGMENT=fast or large

In these configurations the heap is created with a fixed size at the beginning of execution. The default size works well in most cases but if it doesn’t a different size can be specified, as discussed in the following section.

GASNet Communication Layer and the Heap 

In contrast to the dynamic heap extension available in the ugni comm layer, when the GASNet comm layer is used with a native substrate for higher network performance, the runtime must know up front the maximum size the heap will grow to during execution.

In these cases the heap is used for all dynamic allocations, including arrays. By default it will occupy as much of the free memory on each compute node as the runtime can acquire, less some small amount to allow for demands from other (system) programs running there. Advanced users may want to make the heap smaller than the default. Programs start more quickly with a smaller heap, and in the unfortunate event that you need to produce core files, those will be written more quickly if the heap is smaller. Specify the heap size using the CHPL_RT_MAX_HEAP_SIZE environment variable, as discussed in ugni Communication Layer and the Heap. But be aware that just as in the CHPL_COMM=ugni case, if you reduce the heap size to less than the amount your program actually needs and then run it, it will terminate prematurely due to not having enough memory.

Note that for CHPL_COMM=gasnet, CHPL_RT_MAX_HEAP_SIZE is synonymous with GASNET_MAX_SEGSIZE, and the former overrides the latter if both are set.

Known Constraints and Bugs 

Our PBS launcher explicitly supports PBS Pro, Moab/Torque, and the NCCS site versions of PBS. It may also work with other versions. If our PBS launcher does not work for you, you can fall back on a more manual launch of your program. For example, supposing the program is compiled to myprogram:
- Launch the myprogram_real binary manually using aprun and your own qsub script or command.
- Use ./myprogram --generate-qsub-script to generate a qsub script. Then edit the generated script and launch the myprogram_real binary manually as above.
Redirecting stdin when executing a Chapel program under PBS/qsub may not work due to limitations of qsub.
For EX and XC systems, there is a known issue with the Cray MPI release that causes some programs to assert and then hang during exit. A workaround is to set the environment variable, MPICH_GNI_DYNAMIC_CONN to disabled. Setting this environment variable affects all MPI programs, so remember to unset it after running your Chapel program.
The amount of memory available to a Chapel program running over GASNet with the aries conduit is allocated at program start up. The default memory segment size may be too high on some platforms, resulting in an internal Chapel error or a GASNet initialization error such as:
```
node 1 log gasnetc_init_segment() at $CHPL_HOME/third-party/gasnet/gasnet-src/aries-conduit/gasnet_aries.c:<line#>: MemRegister segment fault 8 at  0x2aab6ae00000 60000000, code GNI_RC_ERROR_RESOURCE
```
If your Chapel program exits with such an error, try setting the environment variable CHPL_RT_MAX_HEAP_SIZE or GASNET_MAX_SEGSIZE to a lower value than the default (say 1G) and re-running your program. For more information, refer to the discussion of CHPL_RT_MAX_HEAP_SIZE above and/or the discussion of GASNET_MAX_SEGSIZE here:
```
$CHPL_HOME/third-party/gasnet/gasnet-src/README
```

To request…	set CHPL_TARGET_COMPILER to…
…the LLVM/clang backend	llvm (default)
…the GNU compiler (gcc)	gnu
…the Clang compiler	clang
…the Intel compiler (icc)	intel