Re: Zimbra capacity

[Steve Hillman <hillman@sfu.ca>]

Forwarding along a couple of replies that ended up in my inbox

----- Forwarded Message -----
From: rob@wayne.edu
To: hillman@sfu.ca
Sent: Thursday, October 29, 2009 3:17:00 PM GMT -08:00 US/Canada Pacific
Subject: 


Agreed that this is very difficult to predict, but here are stats from our servers:

  http://eiprod.wayne.edu/zimbrastats/

We have ~60,000 users spread over 6 mailstores (~10,000 users per mailstore)

 - RHEL 64-bit
 - 8 2.5Ghz XEON cores per mailstore (2 physical procs)
 - 16GB ram/mailstore

Hope this helps!
-Rob

-- 

Rob Thompson, Systems Analyst Enterprise Applications 
Computing & Information Technology 
Wayne State University 
313-577-5645 

Public Key: http://pgp.wayne.edu/rob.key 

----- "doug curtis" <doug.curtis@oit.gatech.edu> wrote:

> Hello all!
> 
> I am trying to determine what the max amount of users/traffic each one
> of my mailstores can handle.
> 
> What sort of load average, soap sessions, cpu usage and I/O TPS are
> you seeing per server?
> 
> Thanks,
> 
> Doug
> 
> -- 
> Doug Curtis
> doug.curtis@oit.gatech.edu
> Georgia Tech OIT/A&I
> 404.385.0390


----- Forwarded Message -----
From: bburtin@zimbra.com
To: hillman@sfu.ca
Sent: Thursday, October 29, 2009 3:16:20 PM GMT -08:00 US/Canada Pacific
Subject: 

At the risk of digging up old issues that you've already resolved... have you tried running with a smaller heap size?  Your GC activity looks spiky.  GC happens once every 30 minutes, at which point the heap size drops from 5GB to 1.5GB.  Maybe if you decrease the heap size, garbage collection would happen more frequently, but wouldn't take as much as 5-7% of your CPU?

----- "Rich Graves" <rgraves@carleton.edu> wrote:

> There's a *lot* of variables that go into this, but 3000 users (1900
> undergrad, rest fac/staff) on dual quad core (X5450) running ZCS
> 5.0.19 on RHEL 4.8 over Compellent FC SAN looks like:
> 
> http://mail.carleton.edu/charts/2009-10-28/
> 
> Incidentally, I tried UseConcMarkSweepGC as suggested at
> Performance_Tuning_Guidelines_for_Large_Deployments, but switched back
> to UseParallelGC after three incidences of severe stalls in the
> kernel, each causing 15 minutes of downtime. All CPU cores were 100%
> pegged in system, unable to process interrupts or keep time... and
> then they came back. I'd assume that the updated kernel in RHEL 5
> and/or the updated Java in ZCS 6 would fix whatever weirdness I've run
> into. Zimbra Support helped me address my error contributing to the
> first outage, but had no ideas on the second and third.
> 
> Has anyone else done serious real-world comparison of UseParallelGC
> versus UseConcMarkSweepGC? In theory, the latter is much better, but
> it seems to have jinxed my server, so I backed off. Server was
> completely stable before trying UseConcMarkSweepGC and has been
> completely stable since.
> -- 
> Rich Graves http://claimid.com/rcgraves
> Carleton.edu Sr UNIX and Security Admin
> CMC135: 507-222-7079 Cell: 952-292-6529