Alvin,

> The performance will also depend on storage configuration. 

I would really emphasize this point.

We ran a series of load tests with an SPSS image (each with 1 vCPU and 4 GB) a 
few months ago, and well before hitting any RAM or CPU limits, we hit a wall 
with very high I/O latency.

Initially, our vmhost datastores used shared SAN storage (an array of 7200 RPM 
SATA disks with RAID6). However, once we moved this to an EqualLogic SAS array 
with 15K RPM and RAID 10, the latency dropped to remarkably low levels. 
Ideally, you would use SSD for the running VMs.

Now that I/O latency stays so low, we have been able to oversubscribe the RAM 
of the VMware cluster by about 30%, which means I can pack more VMs onto a host 
than one might think, given the available RAM. Even so, the *actual* memory 
usage of the system when all VMs are running concurrent SPSS analyses typically 
remains between 10% and 25% of the hosts' physical memory. At this point, the 
main resource constraint for us is CPU, which can be addressed by simply 
scaling out.

I would note, however, that in our tests, the VMs were not running anything 
akin to principal component analysis, and so depending on what your users plan 
to do, you may have different resource needs.

Aaron


--
Aaron Coburn
Systems Administrator and Programmer
Academic Technology Services, Amherst College
[email protected]






> 
> 
> On Apr 15, 2013, at 15:28 , Henry Schaffer <[email protected]>
> wrote:
> 
>> On Mon, Apr 15, 2013 at 3:22 PM, alvin allen <[email protected]> 
>> wrote:
>>> Our initial deploy of vcl will be used to run images containing spss and few
>>> other applications.
>>> 
>>> It will be running on a H22 blade with 49000 MB of ram and a Intel (r)
>>> Xeon(R) CPU X5560 @ 2.8GHz with 2.
>>> 
>>> What would would be a rough guess at how many simultaneous users this can
>>> handle?
>> 
>> It will depend greatly on what the users are doing.  I.e. there is a
>> huge difference between running large scale factor analyses (or PCA,
>> which has a similar computational load) or small data descriptive
>> statistics.
>> 
>> For the light weight loads, my *guess* is well above 10 simultaneous  users.
>> 
>> --henry schaffer
>>> 
>>> Thanks in advance,
>>> Alvin
>> 
> 
> 
> 
> --
> Thank you,
> 
> Dmitri Chebotarov
> VCL Sys Eng, Engineering & Architectural Support, TSD - Ent Servers & 
> Messaging
> 223 Aquia Building, Ffx, MSN: 1B5
> Phone: (703) 993-6175 | Fax: (703) 993-3404
> 
> 
> 

Reply via email to