Dear Marawan,
To follow up on Jason's comments I would like to make it clear that
Scaling is a completely pointless and useless metric. You should compare
the performance you are seeing on the new cluster against the old cluster.
That is the metric that matters in this case. Is the new system slower
than the old one?
All the best
Ross
On 9/19/12 10:50 PM, "marawan hussain" <marawanhussain.yahoo.com> wrote:
>Hi Jason,
>Thanks for the reply...What is weird for me is that the Cluster is a very
>new Tesla cluster and the connection of the infiniband type and the
>compilation was carried out using the new mvapich2 library. I'm aware of
>all this issues like NVT to NPT and PMEMD versus Sander..Although the
>previous cluster was old and the connection was ethernet, the scaling was
>much better....If you have some clues about the most potential possible
>source of errors in such cases i will appreciate your help...
>
>Cheers
>
>
>________________________________
> From: Jason Swails <jason.swails.gmail.com>
>To: AMBER Mailing List <amber.ambermd.org>
>Sent: Thursday, 20 September 2012 1:59 PM
>Subject: Re: [AMBER] performance
>
>On Wed, Sep 19, 2012 at 11:43 PM, marawan hussain
><marawanhussain.yahoo.com>wrote:
>
>> Dear AMBER usesrs,
>> I tried to monitor the performance with different libraries and number
>>of
>> processors, the results are in the follwoing table:
>> Number of processor Performance(ns/day) MPI Library
>> 2 0.15
>> mvapich2
>> 4 0.21
>> mvapich2
>> 8 0.28
>> mvapich2
>> 16 0.20
>> mvapich2
>> 32 0.03
>> mvapich2
>>
>>
>>
>> Could someone comment on this
>>
>
>/* http://ambermd.org/amber10.bench1.html (similar performance for pmemd
>12) */
>
>Joking aside, many things affect performance -- interconnect, compilers,
>your computer system, your simulated system, sander vs. pmemd (pmemd is
>much better), PME vs. GB (GB scales much better), NTP vs. NVT vs. NVE (NVE
>is fastest), ig=-1 for Langevin dynamics vs. ig>0 (ig=-1 scales better for
>langevin dynamics), and so on (there are a lot more).
>
>If you can't get your calculations to scale similar to what is found on
>the
>benchmark site, I would probably blame poor interconnect first (note it's
>not just the speed of the interconnect, but also the interconnect topology
>that plays a role in modulating parallel scaling). Unfortunately, if your
>cluster has a bottleneck that isn't present in the benchmark test systems,
>there's no getting around that unless you're willing to replace that
>hardware (and there are way to many possibilities for us to benchmark even
>a representative fraction of different cluster configurations).
>
>HTH,
>Jason
>
>--
>Jason M. Swails
>Quantum Theory Project,
>University of Florida
>Ph.D. Candidate
>352-392-4032
>_______________________________________________
>AMBER mailing list
>AMBER.ambermd.org
>http://lists.ambermd.org/mailman/listinfo/amber
>_______________________________________________
>AMBER mailing list
>AMBER.ambermd.org
>http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Sep 20 2012 - 10:00:03 PDT
