[gmx-developers] MPI_sendrecv and specific timing
Nathan Moore
nmoore at physics.umn.edu
Wed Jul 13 21:33:16 CEST 2005
Fellows,
So I've included the revised version of move_f and move_x from the old
source revision (1.8) of mvxvf.c. I notice a slight speedup in the dpcc
benchmark, but signifigant time remains in the MPI Sendrecv command -
(about 10% of execution wall time). The only other place in the source I
can find MPI_Sendrecv is in stat.c's moveit function, used via #define
with the MX( macro.
Should moveit really be eating up 10% of execution?
regards,
Nathan Moore
timing, dpcc, 5000 timesteps, np=64, in seconds, median processor
statistics below:
version 3.2
MPI Routine #calls avg. bytes time(sec)
-----------------------------------------------------------------
MPI_Comm_size 1 0.0 0.000
MPI_Comm_rank 1 0.0 0.000
MPI_Send 749 27628.6 0.214
MPI_Isend 693883 32055.4 1.240
MPI_Recv 749 27628.6 23.499
MPI_Irecv 693883 32055.4 0.562
MPI_Sendrecv 320190 45696.0 195.866
MPI_Wait 2028146 0.0 238.088
v 3.2 with collective move_f and move_x
MPI Routine #calls avg. bytes time(sec)
-----------------------------------------------------------------
MPI_Comm_size 1 0.0 0.000
MPI_Comm_rank 1 0.0 0.000
MPI_Send 749 27628.6 0.218
MPI_Isend 693883 32055.4 1.301
MPI_Recv 749 27628.6 14.872
MPI_Irecv 693883 32055.4 0.484
MPI_Sendrecv 160158 45696.0 108.985
MPI_Wait 1708082 0.0 245.779
MPI_Allgatherv 5001 45696.0 68.676
More information about the gromacs.org_gmx-developers
mailing list