Kovacs, Corey J. wrote:
It's been a while since I've worked on the following problem but here I am
at it again.
I have a three node system running RHEL3 update 5 (kernel 2.4.21-32) with
GFS-6.0.2.20-1. All three nodes are running as both lock managers and
filesystem clients. When sending thousands of files to the cluster
(on the order of 1/2 terrabyte of 50k files) target nodes will run
out of memory and refuse to fork. Interestingly enough this condition
does not cause the cluster to fence the node, rather it things everything
is "OK". The effect of course is that the fs is not accessable cuz the
cluster is waiting to hear back from the node in question.
I set the high water mark to 10000 (I know that's low, but I wanted to see
the effect)
and the system seemed to be trying to free locks every ten seconds as it
should but
simply could not keep up with the file xfer going in.
By the time a node finally locks up there are over 300K of locks in use.
There is
only a small % diff between the locks reported and the inodes in the
filesystem. If
I interperet this correctly, it simply meand that for almost all the files I
was able to xfer, there is an existing lock being used. Also, mem usage for
lock_gulmd was at 85M+.
When we started logging things it was at 30M+ rising about 3-400k per min.
By reading your description, we may have addressed this issue in RHEL 3
Update 7 which is currently in QA stage.
Note that GFS caches file locks for optimization purpose. The lock is
one-to-one corresponding to VFS inode. BTW, what is the size of your
memory ? Since linux OS itself normally desn't release inode unless
under memory pressure so GFS lock lingers around. In your case, when the
memory pressure starts to show, the inode releasing from base kernel
could be too late for you. Also you have combined lock servers with GFS
nodes - this definitely makes the situation worse. The current solution
is piggybacking the inode purge code in one of our kernel daemons which
wakes up in a tunable interval to purge a tunable percentage of inode.
There are a set of test RPMs available if you're willing to try out. Let
us know.
-- Wendy
--
Linux-cluster@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/linux-cluster
