Charm++ parallel programming system

["Sébastien Boisvert" <sebhtml AT gmail.com>]

Hello,

According to the benchmark 

    http://www.hpcadvisorycouncil.com/pdf/NAMD_analysis.pdf

CHARM++ utilizes these MPI functions:

  - MPI_Get_count
  - MPI_Iprobe
  - MPI_Isend
  - MPI_Recv
  - MPI_Test
  - MPI_Wtime


With most of the time being spent in MPI_Iprobe.


>From the CHARM++ source code file

    git-clones/charm/src/arch/mpi/machine.c

In MPISendOneMsg(), I understand that messages are sent with MPI_Isend and 
that MPI_Test is used to see whether or not buffers can be safely
reused.

In PumpMsgs(), it seems to me that there are at least two code paths:

First, the code attempts to use non-blocking reception with a bunch 
of MPI_Irecv + MPI_Testany + MPI_Get_count.

If that does not work, the code probes for incoming messages, and read one
if any with MPI_Iprobe + MPI_Get_count + MPI_Recv.

But in the benchmarks above on NAMD, they report only the code path with 
MPI_Iprobe, MPI_Get_count and MPI_Recv and none for the code path with
MPI_Testany, MPI_Irecv and MPI_Get_count.

It seems to me that non-blocking communication should be better as the MPI
library can put the incoming data directly in the buffer provided by the user
whereas with MPI_Iprobe it is not necessarily the case.


So what's going on ? Is it because NAMD was compiled with special options
for the benchmarks ?

Thank you.