Choosing the ARMCI Library¶
Overview¶
The Global Arrays parallel environment relies upon a one-sided communication runtime system. There are at least three options currently available:
- ARMCI - This is the original one-sided library developed with Global Arrays by PNNL. It supports the widest range of platforms. See http://hpc.pnl.gov/globalarrays/support.shtml for details.
- ComEx - This library is being developed by PNNL to replace ARMCI. It supports the fewer platforms than ARMCI but may be more robust on modern ones like InfiniBand. See http://hpc.pnl.gov/globalarrays/support.shtml for details.
- ARMCI-MPI - This library is a completely separate implementation of the ARMCI interface by Argonne and Intel. It uses the one-sided communication features of MPI, rather than a platform-specific conduit. It currents provides both MPI-2 and MPI-3 implementations; the MPI-3 one is supported by nearly all platforms. See https://github.com/pmodels/mpich/blob/main/doc/wiki/Index.md for details.
By default, Global Arrays will choose ARMCI or ComEx, based upon the environment variables selected by the user (see GA documentation for details). When a native implementation is available and works reliably, this is the best option for the NWChem user. However, in cases where an native implementation of ARMCI is not available or is not reliable, the user should consider using one of the MPI-based implementations.
There are many different ways to use MPI as the communication runtime of Global Arrays:
| ARMCI_NETWORK | Result | Notes |
|---|---|---|
MPI-PR |
ARMCI with progress rank | Recommended, except on Blue Gene/Q. |
MPI-PT |
ARMCI with progress thread | Appropriate for Blue Gene/Q. |
MPI-MT |
ARMCI over multi-threaded MPI | Do not use Open-MPI 1.x. |
MPI-TS |
ARMCI over MPI without data server | |
MPI-SPAWN |
ARMCI using MPI dynamic processes | Requires MPI_Comm_spawn support. |
ARMCI |
Uses ARMCI-MPI | See the ARMCI-MPI NWChem page for details. |
| (please use mpi3rma branch) | requires EXTERNAL_ARMCI_PATH |
It is difficult to provide complete guidance to the user as to which option to choose. However, we observe the following:
ARMCI_NETWORK=MPI-PRis stable and performs well on many platforms (including Cray XC platforms, e.g. NERSC Cori). This port will use one processes on each node for communication, therefore subtracting one process (again on each node) for NWChem. Therefore, when executing on a single node (i.e. the case of desktop execution) you would need to ask for n+1 processes; in other words, a serial execution would require the following mpirun invocationmpirun -np 2 ...- On Intel True Scale and Omni Path systems,
MPI-PRis more reliable thanOPENIBorMPI-SPAWN.ARMCI-MPIwith Casper and Intel MPI is also recommended. See this page for details. Contact Jeff Hammond for assistance. - On IBM Blue Gene/Q, one must use the mpi2rma branch of
ARMCI-MPIbecause MPI-3 is not fully supported. On all other platforms, theARMCI-MPIbranch mpi3rma is recommended. - The performance of
ARMCI-MPIis greatly enhanced by Casper. See this link for design details and this page for instructions on how to use it.
When using ARMCI-MPI, please make sure to use the most recent version of MPI (MPICH 3.2+, Cray MPI 7.2+, MVAPICH2 2.0+, Intel MPI 5.1+, Open-MPI 2.0+). Older versions of MPI are known to have bugs in the MPI-3 RMA features that affect the correctness of NWChem.
Support¶
Global Arrays, ARMCI and ComEx are developed and supported by PNNL. The user list for support is hpctools@googlegroups.com.
ARMCI-MPI is developed by Argonne and Intel. All ARMCI-MPI questions should be directed to armci-discuss@lists.mpich.org. ARMCI-MPI is not an Intel product.
Automated installation of ARMCI-MPI¶
If you wish to use ARMCI-MPI, a script is available to automatically install it:
cd $NWCHEM_TOP/tools && ./install-armci-mpi