Hi Matthieu, We have to quit meeting like this :)
There were a couple server-side memory leaks fixed (I can't recall if they were specifically tied to write calls) since the 2.8.2 release. I would recommend going to the latest OrangeFS release; there are several bugs related to cancellation (which can occur with congestion associated with load testing) that are fixed in that release. If that isn't an option or you still see the behavior after upgrading, you can run the pvfs2-server process inside valgrind. After the server process stops responding (or it eats a significant amount of memory) kill the process and it should provide some output related to lost memory. If you can provide that it'll give us a starting point. Sorry for your troubles, Michael On Mon, Mar 28, 2011 at 5:36 PM, Matthieu Dorier < [email protected]> wrote: > Hello, > > I'm (again) encountering a bug with PVFS (2.8.2): I use it with TCP on 16 > nodes, an MPI application is running on 38 other nodes and periodically > output some data using MPI-IO. The monitoring system of my PVFS nodes shows > that the memory cached increases at each write, until reaching 8GB (out of > 24GB available on the nodes). At this point the servers stop answering to > write requests, the application blocs in IO and eventually timeout. > I tried with IP over IB and with Ethernet, both show the same behavior. Any > idea where it comes from? > > Thank you, > > Matthieu > > -- > Matthieu Dorier > ENS Cachan, Brittany (Computer Science dpt.) > IRISA Rennes, Office E324 > http://perso.eleves.bretagne.ens-cachan.fr/~mdori307/wiki/ > > _______________________________________________ > Pvfs2-users mailing list > [email protected] > http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users > >
_______________________________________________ Pvfs2-users mailing list [email protected] http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
