Joe Landman wrote:
Tom O'Connor wrote:
Hi List,
We currently have a very irritating problem with Centos 5.3 x86_64
running on a Dell Poweredge SC1435. The problem is this: We are
experiencing frequent kernel panics while using glusterfs and Fuse.
Across the cluster of servers, we are experiencing roughly 1 panic
every 1-2 days. This wasn't a problem with earlier servers where we
used Fedora 6.
Here's a kernel panic screenshot:
http://imagehost.gr/images/c5ad2d5jzgpgoq91v24y.png
Here's some general info:
Linux server6 2.6.18-128.1.10.el5 #1 SMP Thu May 7 10:35:59 EDT 2009
x86_64 x86_64 x86_64 GNU/Linux
Using:
fuse-2.7.4-8_10.el5
fuse-kmdl-2.6.18-128.1.10.el5-2.7.4-8_10.el5
fuse-libs-2.7.4-8_10.el5
glusterfs-common-2.0.1-1.el5
glusterfs-client-2.0.1-1.el5
glusterfs-server-2.0.1-1.el5
I've straced glusterfs while it dies, and there's nothing seriously
spurious, just it stops working as soon as the kernel locks up.
This is odd in that I'd expect Centos to be more stable than Fedora.
We just helped another group with something like this, and the
solution was to move from 2.0.1 to 2.0.6 (and a few other things as
well).
A little background, Gluster is used to share some directories which
are used by apache to serve files from.
I've managed to replicate the live environment inside a virtual
machine, and also to replicate the kernel panic by loading the
virtual machine's apache with ApacheBench, at as few as 3 concurrent
requests, the kernel locks up.
However, i have been unable to reproduce this exact behavior on the
live cluster, and have tried up to 10,000 concurrent requests which
max out the network more than anything.
I don't understand ... its crashing on the live server, but not when
you run ApacheBench?
Yep, It's crashing under "normal" usage (ie, i can't see any traffic
spikes at the time of the crash), but if i try and max it out and really
hammer it with ab, there's no problem at all, slow, but works, and
doesn't fall over.
I've tried latest versions of gluster and fuse from development
snapshots and stable releases, I've tried patched versions of fuse
Which snapshots and which stable releases? 2.0.6 is/was stable at
last check. I wouldn't suggest running production on the dev releases.
I tested primarily on the Virtual Machine, and tried the latest from
centos repositories, i think that's 2.0.1, and the one from doing a
direct git clone on the dev repo, so effectively a nightly build from
about 5 days ago. (it worked, but still exhibited the same behavior).
What fuse's have you tried, what releases, and do they all exhibit the
same behavior? If so, its probably not glusterfs.
Tried the fuse from latest release from the sourceforge page, and the
one in centos 5.3, 2.7.4. Also the latest patched one from Z-research
that is apparently better for gluster.
released by Gluster. Nothing seems to improve this problem.
Another problem we see is the Redhat kernel. We haven't seen it
stable under heavy loads in our storage systems. We have been
building and supporting late model kernels for our units which do
handle load without crashing.
I'm primarily an ubuntu person, and I'll probably test this scenario
against karmic koala (9.10) where gluster is 2.0.1, currently, Jaunty,
only has 1.3.x and it's not easily compaitble with our current
configuration.
If anyone has any ideas for further debugging, or other routes for
support. I'm running out of ideas.
Thanks in advance
Tom O'Connor
--
Tom O'Connor
--------------------------
Assanka: Every possibility
w: http://www.assanka.net/
t: 0870 085 2038
f: 0871 433 0919
e: tom.oconnor@xxxxxxxxxxx
