James,
I have a CentOS 5.2 cluster where I would see the same nfs errors under
certain conditions. If I did anything that introduced latency to my gfs
operations on the node that served nfs, the nfs threads couldn't service
requests faster than they came in from clients. Eventually my nfs
threads would all be busy and start dropping nfs requests. I kept an eye
on my nfsd thread utilization (/proc/net/rpc/nfsd) and kept bumping up
the number of threads until they could handle all the requests while the
gfs had a higher latency.
In my case, I had EMC Networker streaming data from my gfs filesystems
to a local scsi tape device on the same node that served nfs. I
eventually separated them onto different nodes.
I'm sure gfs_grow would slow down your gfs enough that your nfs threads
couldn't keep up. NFS on gfs seems to be very latency sensitive. I have
a quick an dirty perl script to generate a historgram image from nfs
thread stats if you are interested.
-Andrew
--
Andrew A. Neuschwander, RHCE
Linux Systems/Software Engineer
College of Forestry and Conservation
The University of Montana
http://www.ntsg.umt.edu
andrew@xxxxxxxxxxxx - 406.243.6310
James Chamberlain wrote:
Hi all,
I'd like to thank Bob Peterson for helping me solve the last problem I
was seeing with my storage cluster. I've got a new one now. A couple
days ago, site ops plugged in a new storage shelf and this triggered
some sort of error in the storage chassis. I was able to sort that out
with gfs_fsck, and have since gotten the new storage recognized by the
cluster. I'd like to make use of this new storage, and it's here that
we run into trouble.
lvextend completed with no trouble, so I ran gfs_grow. gfs_grow has
been running for over an hour now and has not progressed past:
[root@s12n01 ~]# gfs_grow /dev/s12/scratch13
FS: Mount Point: /scratch13
FS: Device: /dev/s12/scratch13
FS: Options: rw,noatime,nodiratime
FS: Size: 4392290302
DEV: Size: 5466032128
Preparing to write new FS information...
The load average on this node has risen from its normal ~30-40 to 513
(the number of nfsd threads, plus one), and the file system has become
slow-to-inaccessible on client nodes. I am seeing messages in my log
files that indicate things like:
Oct 8 16:26:00 s12n01 kernel: rpc-srv/tcp: nfsd: got error -104 when
sending 140 bytes - shutting down socket
Oct 8 16:26:00 s12n01 last message repeated 4 times
Oct 8 16:26:00 s12n01 kernel: nfsd: peername failed (err 107)!
Oct 8 16:26:58 s12n01 kernel: nfsd: peername failed (err 107)!
Oct 8 16:27:56 s12n01 last message repeated 2 times
Oct 8 16:27:56 s12n01 kernel: rpc-srv/tcp: nfsd: got error -104 when
sending 140 bytes - shutting down socket
Oct 8 16:27:56 s12n01 kernel: rpc-srv/tcp: nfsd: got error -104 when
sending 140 bytes - shutting down socket
Oct 8 16:27:56 s12n01 kernel: nfsd: peername failed (err 107)!
Oct 8 16:27:56 s12n01 kernel: rpc-srv/tcp: nfsd: got error -104 when
sending 140 bytes - shutting down socket
Oct 8 16:27:56 s12n01 kernel: rpc-srv/tcp: nfsd: got error -104 when
sending 140 bytes - shutting down socket
Oct 8 16:27:56 s12n01 kernel: nfsd: peername failed (err 107)!
Oct 8 16:28:34 s12n01 last message repeated 2 times
Oct 8 16:30:29 s12n01 last message repeated 2 times
I was seeing similar messages this morning, but those went away when I
mounted this file system on another node in the cluster, turned on
statfs_fast, and then moved the service to that node. I'm not sure what
to do about it given that gfs_grow is running. Is this something anyone
else has seen? Does anyone know what to do about this? Do I have any
option other than to wait until gfs_grow is done? Given my recent
experiences (see "lm_dlm_cancel" in the list archives), I'm very
hesitant to hit ^C on this gfs_grow. I'm running CentOS 4 for x86-64,
kernel 2.6.9-67.0.20.ELsmp.
Thanks,
James
--
Linux-cluster mailing list
Linux-cluster@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/linux-cluster
--
Linux-cluster mailing list
Linux-cluster@xxxxxxxxxx
https://www.redhat.com/mailman/listinfo/linux-cluster
