Hi, Yesterday I ran into the situation that PVFS2 I/O performance varied quite a lot using gm. I guess the performance variance is a result of package resents and package drops. (~80000 packages for the test, but 40000 package resents on a node). Using only benchmark results with small package loss/resents the performance looks quite good up to 14 / clients and servers.
Attached you will find a simple skript which uses the gm_counter userspace tool to determine the perf counter differences. It is quite simple to use. If you run a mpi program then you can simply run get_gm_var.sh before and after the execution of the program with mpiexec on the same nodes. The script then prints for each counter value node name and the difference of the counters. Of course you can adapt it for your use, it just shows that it is quite easy and interesting to take network behavior into account... I'm curious if somebody else got similar problems with PVFS2 over gm... Best regards, Julian
get_gm_var.sh
Description: application/shellscript
_______________________________________________ Pvfs2-users mailing list [email protected] http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
