The SGI/Cray Research Origin2000 is a Nonuniform Access Memory Access (NUMA)
shared memory architecture. We looked at the protocol sensitivities of
two different message-passing libraries: MPI and an application-specific
library built using the SHMEM one-sided communication operations
get and put.
The MPI and SHMEM communication protocols are not directly
comparable, but some general conclusions can be drawn from the protocol
sensitivity experiments. Similarities between the MPI and SHMEM results are
as follows:
The general rules of thumb (ring algorithms being similar, transpose and
halfsum algorithms being similar, dff and exchsum being
dissimilar) are the same.
The transpose algorithms show less difference between the median and
maximum performance variation statistics than do the distributed algorithms.
The differences include the following:
Overlap techniques are generally useful for SHMEM, but have only limited
utility for MPI implementations.
For MPI, Experiments B and C tend to show the same or more performance
variation than Experiment A. For SHMEM, Experiments B and C tend to show less
variation than Experiment A.
Patrick H. Worley / (
worleyph@ornl.gov)
PSTSWM Performance Page