[Linux-cluster] very poor rm performance with gfs2 (el6)

[Bob Peterson]

----- Original Message -----
| Hello
| 
| I'm currently testing gfs2 under EL6 (2.6.32-71.18.1.el6.x86_64) 11
| nodes over gigabit iscsi (target is IET on EL5)
| Application is a compute cluster with lot of files of various sizes.
| Application need close to write coherency and cache coherency.
| 
| I can't tell yet about performance. It looks not bad, iscsi target
| shows a bandwidth of about 80Mbytes/s (mostly reads during compute). I
| have to compare to NFS to tell if it is good.
| 
| Major problem: I noticed very very poor rm performance, down to 1 to 2
| files erased per second in a rm -rf.
| Some of these files are not very big: a few kilo bytes. Cluster is
| idle during rm. mkfs.gfs2 was made just before compute test was made.
(snip)
| Please advise if there is a solution for this.
| 
| Thanks

Hi,

I've been working on a set of patches to improve gfs2 unlink (aka rm)
performance.  Here is the bug I'm using to track my work:

https://bugzilla.redhat.com/show_bug.cgi?id=681902

I don't know if the bug is public or if anyone has the ability
to see it, but I can't change that (sorry).  I currently have a
"proof of concept" patch that speeds up unlink by 19% in the
tests I'm doing, which is "rm -fR /mnt/gfs2/*".

It also speeds up "ls -lR *" by 42%.  Of course, mileage will vary
depending on what you're doing.  (No guarantees, etc.)  The patch
still has some problems and still needs some work.  At this point
it's just "proof of concept" so don't get your hopes up.  (When I
straighten out the patch, the performance may not be as good.)
It will take a while before anything is available in an official
release.

Regards,

Bob Peterson
Red Hat File Systems