Re: MPICH 1.2 performance
Hi,
In Thu, 27 Jul 2000 22:01:57 +0530 viral@seul.org cum veritate scripsit :
> I wrote a simple program to find out the messages / sec that MPICH could
> do.
> I used the deb of 1.2 as was posted on this list some time back.
> I attach the program hereby.
> What was surprising is that I got only around 100 messages / sec where each
> message is about 350 bytes send, 350 bytes recv.
> LAM gave me something like 2000.
>
> I have compiled MPI myself and run the program, but its given the same
> results, 100 messages/sec.
> However, if i compile with -mpilog, i see 166 messages/sec, which i thought
> was weird behaviour. All this was replicable.
I found this to be interesting, and I've tried it out.
The machines I have used to test it were roughly...
(No, I don't have a SPARC cluster at home)
Pentium III 500MHz 200Mb memory
80486 SX/33 40Mb memory
linked with 10base-T ethernet with a non-switching hub, and only these
machines active.
Results were :
$ mpicc performance_mpi.c -o performance_mpi
$ mpirun -np 2 performance_mpi
50000 packets transmitted to and fro in 213 seconds
No. of packets per second: 234
$ mpicc -mpilog performance_mpi.c -o performance_mpi
$ mpirun -np 2 performance_mpi
50000 packets transmitted to and fro in 453 seconds
No. of packets per second: 110
Writing logfile.
$ mpicc -mpianim performance_mpi.c -o performance_mpi -lm -L/usr/X11R6/lib -lXt
$mpirun -np 2 performance_mpi -display=:0
50000 packets transmitted to and fro in 808 seconds
No. of packets per second: 61
NOTE: I seem to have to do "xhost +servername" to make it work. If anyone knows of
any way better than to do that, then this animation thing might be quite handy.
I couldn't reproduce it. What might be the problem ?
I haven't yet looked into LAM.
Does it work as expected, including update-alternatives ?
# i.e. should mpich provide "mpi" ?
regards,
junichi
--
University: ti0113@mail4.doshisha.ac.jp Netfort: dancer@netfort.gr.jp
dancer, a.k.a. Junichi Uekawa http://www.netfort.gr.jp/~dancer
Dept. of Knowledge Engineering and Computer Science, Doshisha University.
... Long Live Free Software, LIBERTAS OMNI VINCIT.
Reply to: