README

#
# Copyright (c) 2018      Los Alamos National Security, LLC.  All rights
#                         reserved.
# $COPYRIGHT$
#
# Additional copyrights may follow
#
# $HEADER$
#

The RMA-MT benchmarks are designed to evaluate the performance of an
implementation of the Message Passing Interface (MPI) Remote Memory
Access (RMA) interface. There are three tests in this benchmarking
suite:

rmamt_bw:

  This benchmark tests the bandwidth of the MPI implementations when
  subjected to many concurrent put or get operations. It prints out
  the effective bandwidth and the message rate for each message size.

rmamt_bibw:

  This benchmark tests the bi-directional bandwidth of the MPI
  implementations when subjected to many concurrent put or get
  operations from both MPI processes. It prints out the effective
  bi-directional bandwidth and the message rates for each message
  size.


rmamt_lat:

  This benchmarks tests the multi-threaded latency of the MPI
  implementation. This is evaluated by summing the total time for each
  thread to do a single put or get operation plus the time for
  synchronization. It prints out the total latency for each message
  size.


Common options for all benchmarks:

 -o,--operation=value      Operation to benchmark: put, get
 -s,--sync=value           Synchronization function to use: lock_all,
                           fence, lock, flush, pscw
 -m, --max-size=value      Maximum aggregate transfer size
 -l, --min-size=value      Minimum aggregate transfer size
 -w,--win-per-thread       Use a different MPI window in each thread
 -i,--iterations=value     Number of iterations to run for each per
                           message size (default: 1000)
 -t,--threads=value        Number of threads to use
 -n,--busy-loop            Run busy loop on receiver
 -x,--bind-threads         Bind threads to unique cores (requires
                           hwloc)
 -z,--sleep-interval=value Sleep interval in ns to use on receiver if
                           using busy loop. Not relevant to rmamt_bibw
                           loop (default: 10000)
 -r,--result=value         Write parsable output to the specified file
                           in CSV format.
 -h,--help                 Print this help message


Configuring and building:


If the desired MPI compiler wrapper is in your path (mpicc, cray cc)
you can simply run:

./configure
make


You can also specify the desired wrapper by setting the CC environment
variable.



Running:


The RMA-MT benchmark suite comes with some basic run scripts in the
scripts directory. These can be modified to run the entire
benchmarking suite in the current allocation using the mpirun
launcher. They can be modified to use any launcher. Note that by
default the scripts bind the bechmark to the socket (unless -x is
specified to bind the worker threads).

The benchmarks can also be run by hand. The minimum arguments required
are -o and -s to specify the operation (put, get) and the
synchronization method (flush, lock, lock_all, pscw). There must be
exactly two processes.

Ex:

mpirun -n 2 --npernode 1 rmamt_bw -o put -s flush -t 32


This will run the bandwidth benchmark with MPI_Put() and
MPI_Win_flush() with 32 threads.