Release66:ARMCI
From NWChem
Choosing the Right ARMCI Library
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/armci/ 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/comex/ 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://wiki.mpich.org/armci-mpi/index.php/Main_Page 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 | Additional options | Result | Notes |
MPI-PR | ARMCI with progress rank | Recommended | |
MPI-PT | ARMCI with progress thread | ||
MPI-MT | ARMCI over multi-threaded MPI | Do not use Open-MPI. | |
MPI-TS | ARMCI over MPI without data server | ||
MPI-SPAWN | ARMCI over MPI using dynamic processes | Requires MPI_Comm_spawn support. | |
ARMCI | EXTERNAL_ARMCI_PATH | Uses ARMCI-MPI (please use mpi3rma branch) | See the ARMCI-MPI NWChem page for details. |
It is difficult to provide complete guidance to the user as to which option to choose. However, we observe the following:
- ARMCI_NETWORK=MPI-PR is stable and performs well on Cray XC platforms (e.g. NERSC Edison).
- On Intel True Scale systems, MPI-MT is more reliable than OPENIB or MPI-SPAWN.
- On IBM Blue Gene/Q, one must use the mpi2rma branch of ARMCI-MPI because MPI-3 is not fully supported. On all other platforms, the ARMCI-MPI branch mpi3rma is recommended.
- The performance of ARMCI-MPI is greatly enhanced 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. 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 Labs. All ARMCI-MPI questions should be directed to armci-discuss@lists.mpich.org. ARMCI-MPI is not an Intel product.