[CP2K:875] Re: performance on Linux Opteron cluster

[LIANG Yunfeng]

Thanks a lot!  I feel that it was so "dangerous" for me if I didn't ask....

Sincerely, Yunfeng

On Mon, Mar 24, 2008 at 2:17 PM, Teodoro Laino <teodor... at gmail.com>
wrote:

> Dear Yunfeng,
>
> I guess you're definitely not aware of what you were running.
> That input file is not there to do performance analysis (there's a
> benchmark directory instead!)
> In few months, when you will have a more solid cp2k background,  you
> will open again that input file and you will realize that
> the file 32H2O-md.inp is highly IO bounded.
> RESTART of the wavefunctions are written at every SCF step and
> restart files at every MD step and since the IO (at least in cp2k)
> are non-parallel operations your code will never scale..
>
> I just modified for you the input file and I have run from 1 to 32
> procs and these are the data:
>
> 1_proc/32H2O-md.out:    - 3294.21
> 2_proc/32H2O-md.out:    - 1898.79  (1.73)
> 4_proc/32H2O-md.out:    - 1038.44  (1.82)
> 8_proc/32H2O-md.out:    -   572.58  (1.81)
> 16_proc/32H2O-md.out:  -   334.94  (1.71)
> 32_proc/32H2O-md.out:  -   242.30  (1.38)
>
> Timings are in seconds and in parentheses the speed w.r.t the same
> job with half number of processor (ideally 2).
> The writing of  the wfn restart file can easily take 1 second each
> time.. and in that run it is going to be written ~500 times.
> This justifies your timings!!!
>
> Keep in mind that these jobs have run on a dual core machine (like
> yours).. it is possible to have ration closer to 2 (?!?) when using
> single core processors.
>
> I can understand the psychology behind the hurries of  testing how
> well a code is scaling.. but...
> to do that you need to know what you are doing.. and know it very
> well.. even with very user friendly codes!
> In attachment all the data, for you and for all the people that in
> the future want something to compare..
>
> Teo
>
>
> >
>
>
>
> On 24 Mar 2008, at 18:44, LIANG Yunfeng wrote:
>
> > Dear All,
> >
> > The analyst help me set up cp2k (with the PGI compiler) in Linux
> > Opteron cluster with Infiniband interconnect, where each node has 2
> > Opteron CPUs (2.40GHz) and 2 GB of memory.
> >
> > He gives the reports as below:
> >
> > Running on 2 processors, this completed the 32H2O-md.inp
> > problem in about 32 minutes....
> >
> > I also ran it on 4 processors, which took a little over 20 minutes.
> > I then ran it on 8 processors and found it took just under 20 minutes.
> > So, the program is not scaling very well as the number of processors
> > is increased.
> >
> > ....
> >
> > I think cp2k is still heavily bounded with OpenMP, so the
> > performance can only be like the above, am I right?
> >
> > Thank you
> > Sincerely, Yunfeng
> >
> > --~--~---------~--~----~------------~-------~--~----~
> > You received this message because you are subscribed to the Google
> > Groups "cp2k" group.
> > To post to this group, send email to cp... at googlegroups.com
> > To unsubscribe from this group, send email to cp2k-
> > unsub... at googlegroups.com
> > For more options, visit this group at http://groups.google.com/
> > group/cp2k?hl=en
> > -~----------~----~----~----~------~----~------~--~---
> >
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20080324/5d49b596/attachment.htm>