On Mon, 15 Feb 2010 00:43:02 +0100 Anton Starikov <ant.starikov@xxxxxxxxx> wrote: > Hi, > > The setup: > is 16-core opteron node, diskless with NFS root, swapless, 64GB of RAM. Operating under OpenSUSE 11.2. With kernel version 2.6.31. Although it isn't vanilla, I think probably more right is to submit this into LKML. > At first, what is the version of kernel you are comparing with ? 2.6.22?(If OpenSuse10) If so, many changes since that.. > The problem: > On this node user run MPI job with 16 processes, local job by using shared memory communication. > At some point this processes are trying to use more memory that available. > Normally, all of them or part of them would be killed by OOM killer, and it use to work for years over many versions of kernel. > > Now, with fresh setup I got something new. OOM tried to kill, but didn't succeed, and even more, brought system in unusable state. All those processes are locked and un-killable. some of other processes are also locked and un-killable/inaccessible. kswapd consume 100% CPU (which I think is expected behavior when there is no free memory). > No free memory obviously, cause all original processes are still in memory. > > I tried to test OOM behavior and it always happens like that now. > > Here I attach full gzipped log of all related information captured by logserver (sent by logserver and netconsole, so it can be partly doubled). Sorry that it is too big, but I didn't know what information can be important. > Anyway, I think it's not appreciated to depend on OOM-Kill on swapless-system. I recommend you to use cgroup "memory" to encapsulate your apps (but please check the performance regression can be seen or not..) Thanks, -Kame -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>