(MPI/MPL Comparison - March 22, 1996 and July 26, 1996)
The IBM SP2 is a distributed-memory parallel
architecture utlizing high-end workstation-class processors
interconnected by an omega switch.
We looked at the protocol sensitivities of two different message-passing
libraries: MPI and the original native message-passing
library MPL.
The MPI and MPL communication protocols used in these experiments are directly
comparable, except that MPL does not have a real ready send. We emulated
this by using the normal send with the ready send handshaking logic,
but it was never competitive with the simpler protocols.
Similarities between the MPI and MPL results are as follows:
With the exception of dfft, the MPI and MPL algorithms show
similar performance variation.
When an optimal MPI protocol does not use ready send, it is often
also an optimal MPL protocol.
The general rules of thumb that srtrans and swtrans are
similar and that exchsum, halfsum, and logtrans are
similar hold for both MPI and MPL.
The differences include the following:
MPI optimal protocols often include ready send and overlap techniques,
and never use MPI_SENDRECV.
MPL optimal protocols often use the native sendrecv, use overlap less often,
and never use the (emulated) ready send protocols.
The simple propotionality relationship between performance variation and
granularity is stronger in the MPI results than in the MPL results.
Patrick H. Worley / (
worleyph@ornl.gov)
PSTSWM Performance Page