[Pw_forum] Insufficient Virtual Memory

[Vo, Trinh]

Dear Axel,

Thanks for clarification.

About the benchmarks, I just simply to see how well is the performance of  the cluster we bought in term of scaling with QE.  I sent some plots to you, but the email did not go thru because of the restriction of the size (larger than 40K).

Currently, I am not happy at the fact that the difference in CPU time and wall time is too large.     When I run a longer job, which took ~2h CPU time long, the wall time was ~7h when I run from the head node, and ~4h when I run from another node, which is not the head node.  The reason why I avoided running from the head node was that other users might use the head node to do their work, causing the long wall time.  Even when I did the test runs (short runs), when no one used the cluster, the difference was still up to 25%, depending on how many nodes I used. I do not know what I should do to improve this difference. According to you, what should I look at to fix the problem. (I also need to read the forum discussion as Lorenzo said in his email).

Thanks,

Trinh

On 2/26/09 4:34 PM, "Axel Kohlmeyer" <akohlmey at cmm.chem.upenn.edu> wrote:

On Thu, 26 Feb 2009, Vo, Trinh wrote:

TV> Dear Axel,
TV>

TV> I am sorry.  I may be out of date.  Is the cp2k code included in QE
TV> package, or in some other codes?  I saw the word "cp2k" in the title

perhaps i was not clear enough. cp2k is a completely different project
and has not much else to do with Q-E outside of the fact that both
have a DFT module and that i and a few others are using both codes.

TV> of the plot, but I thought that it might be the same cp code of QE.

no. i already wrote that. in fact, "there is no CP in CP2k".
i suggested this phase as a subtitle to the project, but people
found it to negative. ;)

TV> Actually, the benchmark plot that I obtained does not look as good
TV> as the posted plot.  I will send it to you in the next email. I need
TV> to make a better plot before I send it.

benchmark in terms of scaling or in terms of problem set size?

there should be a post in the cp2k google group referring to the
benchmark graph and describing the exact system setup.

cheers,
   axel.

TV>
TV> Thanks,
TV>
TV> Trinh
TV>
TV>
TV> TV> tried to repeat all the calculations to compare, but I could not run
TV> TV> that case.
TV>
TV> this graph was done with a _very_ different code, cp2k, that
TV> uses a method (quickstep) different from the CP in cp.x and has
TV> therefore different memory requirements and system size scaling.
TV> the README of example21 explicitly warns you about running out
TV> of memory. give the current difference in speed between hard
TV> drives and main memory, every calculation where you run into
TV> using swap is not worth the effort (unlike 10 years ago).
TV>
TV> just for reference: because of the difference in methods cp.x
TV> should scale better with the number of processors, whereas
TV> cp2k should scale better with the problem set size than cp.x
TV> with the same number of processors.
TV>
TV> that being said, cp2k demonstrates the problem of overloading
TV> memory and communication channels very nicely and thus the
TV> graphs are a nice point of reference to showcase how bad it
TV> can get and that the not so obvious choice of not using all
TV> cpus actually gives you the better performance.
TV>
TV> cheers,
TV>    axel.
TV>
TV> TV>
TV> TV> Thank you,
TV> TV>
TV> TV> Trinh
TV> TV>
TV> TV>
TV> TV>
TV> TV> -----Original Message-----
TV> TV> From: pw_forum-bounces at pwscf.org [mailto:pw_forum-bounces at pwscf.org] On Behalf Of Axel Kohlmeyer
TV> TV> Sent: Thursday, February 26, 2009 1:04 PM
TV> TV> To: PWSCF Forum
TV> TV> Subject: Re: [Pw_forum] Insufficient Virtual Memory
TV> TV>
TV> TV> On Thu, 26 Feb 2009, Vo, Trinh wrote:
TV> TV>
TV> TV> TV> Dear PWSCF users,
TV> TV> TV>
TV> TV> TV> I was running a benchmark (using example21 in QE) for my computer cluster (total 18 nodes) to compare the results with those posted on web link that was mentioned by Axel in some previous forum emails..
TV> TV> TV>
TV> TV> TV> When I performed the calculation for the case of 512 H2O with 4cores/node, the job crashed with the error:
TV> TV> TV>
TV> TV> TV>  "forrtl: severe (41): insufficient virtual memory".
TV> TV> TV>
TV> TV>
TV> TV> TV> I do not know how to deal with this.  Any help or suggestion is
TV> TV> TV> appreciated.  For other cases, the runs were fine.
TV> TV>
TV> TV> have you _read_ the README file in that directory???
TV> TV>
TV> TV> axel.
TV> TV>
TV> TV>
TV> TV> TV>
TV> TV> TV> Below are the specifications of my cluster:
TV> TV> TV>
TV> TV> TV> Intel CPU Xeon X5450 Quad Core 3.0GHz with 12M cache 133MHz FSB
TV> TV> TV> DDRII667 4GB RAM FBD ECC (D-667D2D4F5/4GB)
TV> TV> TV>
TV> TV> TV> For compilation, I used the following:
TV> TV> TV> QE ver. 4.0.3
TV> TV> TV> ifort 10.1015, MKL  10.0.3.020, impi 3.1
TV> TV> TV> internal fftw
TV> TV> TV>
TV> TV> TV> I also attach here the output file.
TV> TV> TV>
TV> TV> TV>
TV> TV> TV> Thank you,
TV> TV> TV>
TV> TV> TV> Trinh Vo
TV> TV> TV>
TV> TV> TV>
TV> TV> TV>
TV> TV> TV>
TV> TV> TV>
TV> TV>
TV> TV>
TV>
TV>

--
=======================================================================
Axel Kohlmeyer   akohlmey at cmm.chem.upenn.edu   http://www.cmm.upenn.edu
   Center for Molecular Modeling   --   University of Pennsylvania
Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323
tel: 1-215-898-1582,  fax: 1-215-573-6233,  office-tel: 1-215-898-5425
=======================================================================
If you make something idiot-proof, the universe creates a better idiot.
_______________________________________________
Pw_forum mailing list
Pw_forum at pwscf.org
http://www.democritos.it/mailman/listinfo/pw_forum

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20090227/2764d071/attachment.html>