Flavio Junior wrote:
On Fri, Jul 3, 2009 at 4:40 PM, Gordan Bobic <gordan@xxxxxxxxxx <mailto:gordan@xxxxxxxxxx>> wrote:
GFS1 doesn't have any more tuning options than GFS2 that I can think of. And besides, in practice, if the performance isn't in the right ball park out of the box, no amount of tweaking will help. Just about the only think that makes a significant difference is the noatime mount option. I wouldn't bother with the rest unless you really need those last few percent.
Well, no problem at al to migrate to GFS1, actually I've already thinked about it, but all those gfs1 tunning options and tests makes me a bit apprehensive.
Noatime helps, but where I've seen some really good performance boosts is in tweaking glock_purge and demote_secs parameters. Of course, always start with a modest setting and start tweaking from there. Playing around with statfs_fast =1, noatime, nodiratime and playing withe glock settings, I've seen a pretty significant jump in performance.
The problem with GFS2 is that it's still a bit buggy, as you've found. But there isn't that much difference in performance between various similar file systems. Sure, GFS2 is faster than GFS1, but it's not an order of magnitude faster.I'll wait a bit more for GFS2 community, if they say that it can't be done I go to GFS1 or even ocfs2 (what is the third option, as I've already a RHCS structure with clvmd).
I've done some GFS vs GFS2 performance benchmarking for a cluster that I will be putting in soon. I've found that GFS1 performance has been much much better than GFS2. As far as I can tell, GFS2 lacks a lot of the tunability that GFS1 has. All the documentation I've seen says that it's supposed to be self-tuning, so there are fewer performance tuning options you have to play with. From my tests, I've had almost a 50% reduction in performance using GFS2.
-- Linux-cluster mailing list Linux-cluster@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/linux-cluster
