[mpich-discuss] Strange behavior of mpd

Bellanca Gaetano gaetano.bellanca at unife.it
Tue Sep 30 11:19:09 CDT 2008


Hi Gus and list,

I thought (hope!) that increasing the dimension of the MPI buffer 
size can produce the same effect ... without writing other code. I 
agree with you that having to manage smaller messages should be 
simpler. However, I was surprised to have the same behavior on two 
very different PCs (the one at home, with a single CPU and 4GB of 
RAM, and the one at the University, with 2 processors and 16GB of RAM).
Moreover, I changed the code (before I had MPI_SEND and MPI_RECV, now 
I have MPI_GATHERV implemented), and I had the same error in the same 
point of the code. This seems to confirm that the dimension of the 
message is the problem.

However, when I discovered that simply killing and restarting mpd, or 
having mpd started by hands after the boot, is a solution to this 
problem, I thought that I can use this as a possible way to overcome it.

Rajeev (thanks Rajeev for your comment!) pointed out that mpd doesn't 
interfere with MPI communication; so I can't explain really what's 
happening. But I observed this behavior.

About the dimension of the message, the computational domain is 
326x118x328 wide, and when partitioned by two becomes 163x118x326 -> 
627284 real values. I have no idea if this could be a too big message.

Maybe I did the same mistake compiling and installing mpich, so that 
it doesn't work correctly (from this point of view; results are 
correct and the same running on 1PE and on #PEs)?

Regards.

G


----------
Gaetano Bellanca - Department of Engineering - University of Ferrara
Via Saragat, 1 - 44100 - Ferrara - ITALY
Voice (VoIP):  +39 0532 974809     Fax:  +39 0532 974870
mailto:gaetano.bellanca at unife.it

----------
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/mpich-discuss/attachments/20080930/74d02890/attachment.htm>


More information about the mpich-discuss mailing list