Re: NAMD running on windows HPC question

From: Axel Kohlmeyer (akohlmey_at_gmail.com)
Date: Tue Sep 20 2011 - 11:27:23 CDT

On Tue, Sep 20, 2011 at 9:46 AM, johan strumpfer <johanstr_at_ks.uiuc.edu> wrote:
> Hi Joseph,
>
> I see, hmm - I'm out of my depth then. I presume you're using the
> Win64-MPI namd build that's available on the website. I wonder if
> there's an ibverbs version that can be built for windows, and if this
> may be the issue. Maybe Jim or Axel can comment.

the important questions is: does the MPI version
actually utilize the infiniband link?

i would suggest to try some standard MPI benchmarks
to measure things like all-to-all bandwidth and latencies
for sufficiently large data sets as a worst case scenario.

i would then compare to a similarly configured linux
machine (or just set your machine up for dual-boot).

axel

> Cheers,
> Johan
>
> ------------------------------------------------------------------------------------------------------
> Johan Strumpfer:         johanstr_at_ks.uiuc.edu        www.ks.uiuc.edu/~johanstr
> Theoretical and Computational Biophysics Group
> 3115 Beckman Institute
> University of Illinois at Urbana-Champaign
> 405 N. Mathews
> Urbana, IL 61801, USA
> ------------------------------------------------------------------------------------------------------
>
>
>
> On Tue, Sep 20, 2011 at 9:34 AM, Chewning, Joseph
> <Joseph.Chewning_at_targacept.com> wrote:
>> Johan,
>>
>> While I would agree, we have an Infiniband 40Gb pipe in between the nodes. That is why I was leaning toward the software side.
>>
>> Thanks,
>> Joseph
>>
>> -----Original Message-----
>> From: johan.ks.uiuc_at_gmail.com [mailto:johan.ks.uiuc_at_gmail.com] On Behalf Of johan strumpfer
>> Sent: Tuesday, September 20, 2011 9:31 AM
>> To: Chewning, Joseph
>> Cc: namd-l_at_ks.uiuc.edu
>> Subject: Re: namd-l: NAMD running on windows HPC question
>>
>> Hi Joseph,
>>
>> It sounds like you have a slow network that is hampering your
>> calculation. Although you can crunch more numbers with more
>> processors, NAMD needs to communicate between all nodes. The faster
>> you process, the faster your network needs to be.
>>
>> Cheers,
>> Johan
>>
>>
>> ------------------------------------------------------------------------------------------------------
>> Johan Strumpfer:         johanstr_at_ks.uiuc.edu        www.ks.uiuc.edu/~johanstr
>> Theoretical and Computational Biophysics Group
>> 3115 Beckman Institute
>> University of Illinois at Urbana-Champaign
>> 405 N. Mathews
>> Urbana, IL 61801, USA
>> ------------------------------------------------------------------------------------------------------
>>
>>
>>
>> On Tue, Sep 20, 2011 at 8:58 AM, Chewning, Joseph
>> <Joseph.Chewning_at_targacept.com> wrote:
>>> Hi All,
>>>
>>>
>>>
>>> I am running into a problem on our new Microsoft HPC system running NAMD. We
>>> have two computer nodes with 32 processors each. If we run the job on one
>>> node the bench marks look fine but if we run the job on both nodes it takes
>>> a lot longer. I am using the following command to run the job:
>>>
>>>
>>>
>>> mpiexec -cores 64 \\HPC-PROD\NAMD\2.8\NAMD2.exe
>>> \\HPC-Prod\Jobs\Test_Job\ch25.conf >
>>> \\hpc-prod\Jobs\Test_Job\test_joe\namd.log
>>>
>>>
>>>
>>> Is the CUDA code ready for testing yet?
>>>
>>>
>>>
>>> Thanks,
>>>
>>>
>>>
>>> Joseph
>>>
>>>
>>>
>>> Confidentiality Notice: This message is private and may contain confidential
>>> and proprietary information. If you have received this message in error,
>>> please notify us and remove it from your system and note that you must not
>>> copy, distribute, or take any action in reliance on it. Any unauthorized use
>>> or disclosure of the contents of this message is not permitted and may be
>>> unlawful.
>>>
>>> TARGACEPT, the Targacept logo, and Building Health, Restoring Independence
>>> are trademarks or service marks of Targacept. Nothing in this communication
>>> shall be construed as granting any license or right to use any trademark or
>>> other intellectual property right of Targacept, including by implication,
>>> estoppels, or otherwise.
>> Confidentiality Notice: This message is private and may contain confidential and proprietary information. If you have received this message in error, please notify us and remove it from your system and note that you must not copy, distribute, or take any action in reliance on it. Any unauthorized use or disclosure of the contents of this message is not permitted and may be unlawful.
>>
>> TARGACEPT, the Targacept logo, and Building Health, Restoring Independence are trademarks or service marks of Targacept. Nothing in this communication shall be construed as granting any license or right to use any trademark or other intellectual property right of Targacept, including by implication, estoppels, or otherwise.
>>
>
>

-- 
Dr. Axel Kohlmeyer
akohlmey_at_gmail.com  http://goo.gl/1wk0
Institute for Computational Molecular Science
Temple University, Philadelphia PA, USA.

This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:57:45 CST