Hi becky I think this is the mopid reuse problem from years past. Basically at high load a mopid on one machine gets recycled and used again before getting invalidated on the receiver side so we ends up with a dupe. I dont recall being able to fix this but we lowered the frequency of its occurrance to a point where the kernel module interface was stable by adding locking logic to the mopid usage in bmi-send. I don't really have access to the code anymore but that's where id recommend starting the search. Even with this though we still observed thee problem in heavy usage over native bmi implementations like netpipe and a port of gamess I made use native calls. On Jul 19, 2011 4:00 PM, "Becky Ligon" <[email protected]> wrote:
_______________________________________________ Pvfs2-users mailing list [email protected] http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
