[mvapich-discuss] Incorrect use of pSync/pWrk in OSU OpenSHMEM benchmarks

[Sayan Ghosh]

In some OSU OpenSHMEM benchmarks, there are back to back collective
function calls using the same pSync/pWrk array, which violates the
OpenSHMEM 1.0 spec :

Page 58:

The pWrk and pSync arrays on all PEs in the active set must not be in use
from a prior
call to a collective OpenSHMEM routine.

Some OpenSHMEM implementations might fail when encountered with something
like (an example from OSU OpenSHMEM benchmark):

    shmem_double_min_to_all(&min_time, &latency, 1, 0, 0, numprocs, pWrk,
pSyncRed);
    shmem_double_max_to_all(&max_time, &latency, 1, 0, 0, numprocs, pWrk,
pSyncRed);
    shmem_double_sum_to_all(&avg_time, &latency, 1, 0, 0, numprocs, pWrk,
pSyncRed);


An easy fix would be to declare 2-D pSync/pWrk arrays/or multiple ones and
use then alternatively:

shmem_double_min_to_all(&min_time, &latency, 1, 0, 0, numprocs, pWrk[0],
pSyncRed[0]);
shmem_double_max_to_all(&max_time, &latency, 1, 0, 0, numprocs, pWrk[1],
pSyncRed[1]);
shmem_double_sum_to_all(&avg_time, &latency, 1, 0, 0, numprocs, pWrk[0],
pSyncRed[0]);

Thanks to Brian Barrett of Sandia Labs for pointing this out to me.

-- 
Thanks,
Sayan
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20131024/fcc3b415/attachment.html>