[Linux-cluster] Performance tunning

[Lon Hohberger]

On Fri, 2005-02-18 at 00:45 +0800, David Teigland wrote:
> On Thu, Feb 17, 2005 at 10:47:55AM -0500, Manuel Bujan wrote:
> > Hello,
> > 
> > Is there anyway to improve the performance of a cman based gfs cluster 
> > installation tunning some of its parameters ?
> > 
> > We are using the dlm style locking mechanism in a two node 100 GB shared 
> > gfs cluster with a DELL powervault 220s storage.
> > We already disabled the atime updates and  the quotas_account but the gain 
> > in performance is not too significant.
> > 
> > Any other hints ?
> 
> Look at /proc/cluster/lock_dlm/drop_count
> 
> It's 50000 by default and increasing it could possibly improve things,
> or disable it altogether by setting to zero.
> e.g.  echo "0" >> /proc/cluster/lock_dlm/drop_count
> 
> You need to change this prior to mounting gfs on each node.  When
> non-zero, GFS/lock_dlm tries to keep the number of locks held locally
> below this level which limits the caching gfs can do.  When zero, no
> limiting is attempted.

Additionally, there are performance implications when using JBODs +
host-RAID controllers.  Basically, each controller's caching has to be
disabled (aka "put in cluster mode") in order for most host-RAID
controllers to operate properly in a clustered environment.  Naturally,
disabling the RAID controllers' cache can hurt I/O performance.

It might be interesting to test a one node configuration with and
without cache disabled and see if there is a significant change in
performance with your tests (still using DLM+GFS, of course).

Out of curiosity, what tests/benchmarks are you using?

-- Lon