The Intel Paragon is a distributed-memory parallel
architecture built around a high-performance 2D grid interconnect.
We looked at the protocol sensitivities of two different message-passing
libraries: MPI and the native message-passing library
NX.
The MPI and NX communication protocols used in these experiments are directly
comparable, except that MPI_BSEND does not have an NX equivalent.
We use the data from the NX experiment A1, so that the partitions and math
libraries are the same for the MPI and the NX data.
Similarities between the MPI and NX results are as follows:
The transpose experiments 2 and 3 have significantly higher performance
variations than transpose experiment 1.
The optimal protocols for logtrans and srtrans are similar
between MPI and NX (but not identical) for each of the separate transpose
experiments and parallel algorithms.
The optimal protocols for exchsum, halfsum,
ringpipe, and ringsum are similar between MPI and NX.
The differences include the following:
The performance variations are slightly larger for MPI than for NX
using the median statistic, and much larger using the maximum statistic.
NX optimal protocols often include ready send, while MPI optimal
protocols never do (generally using synchronous send instead).
Patrick H. Worley / (
worleyph@ornl.gov)
PSTSWM Performance Page