[Pw_forum] Re: Re: Re: Woodcrest vs Opteron performance in pwscf calc. (Huiqun Zhou)
hqzhou at nju.edu.cn
Mon Aug 7 11:09:59 CEST 2006
Here I report my test result on a completely clean woodcrest node:
1 cores : 3m54s (3m53.51s)
2 cores: 2m12s (2m10.61s)
4 cores: 1m22s (1m17.38s)
Comparing the results of this test and that in my previous post, we can
(1) Setting OMP_NUM_THREADS to 1 or not has no influence on the speed.
This may be due to the fact that Intel BLAS and LAPACK libraries assume
that it's equal to 1 while the Scalapack and sparse solver in MKL library
(2) It seems that the memory contention is significant:
1 core : 3m54s (clean) vs. 4m44s (loaded on other 2 cores)
2 cores: 2m12s (clean vs. 2m52s (loaed on other 2 cores)
----- Original Message -----
From: "Huiqun Zhou" <hqzhou at nju.edu.cn>
To: <pw_forum at pwscf.org>
Sent: Saturday, August 05, 2006 3:55 PM
Subject: Re: [Pw_forum] Re: Re: Re: Woodcrest vs Opteron performance in
pwscf calc. (Huiqun Zhou)
> Cesar and list-users,
> I'm sorry for my fault. You're right. As there is only 2 cores on one of
> my woodcrest nodes are spare right now, I re-run the tests on them and
> the opteron node.
> Here are the results:
> (1) Opteron 285 (2.6 GHz)
> Nearly same as my previous reported numbers.
> (2) Woodcrest (2.66 GHz)
> 1 core : 4m44s (4m42.74s)
> 2 cores: 2m52s (2m49.49s)
> The results were got while another pwscf job was running on other
> two cores.
> I'll run the same test on the node next week (maybe) when it becomes
> totally clean and spare. So, we can get an impression of how memory
> contention would influence the speed as Axel prompted.
> Have a nice weekend!
> Huiqun Zhou
> ----- Original Message -----
> From: <cesards at msi.umn.edu>
> To: <pw_forum at pwscf.org>
> Sent: Saturday, August 05, 2006 6:04 AM
> Subject: [Pw_forum] Re: Re: Re: Woodcrest vs Opteron performance in pwscf
> calc. (Huiqun Zhou)
>> Dear Huiqun Zhou,
>> In your job there is a line
>> Should be
>> I don't know how many threads your system defaults to, but if it is more
>> than one, your tests are flowed.
>> Cesar R.S. da Silva
>> Pw_forum mailing list
>> Pw_forum at pwscf.org
> Pw_forum mailing list
> Pw_forum at pwscf.org
More information about the users