[gmx-users] linux kernel 2.4 vs 2.6

[David L. Bostick]

Dear all,

I have done some testing on the two different Xeon clusters.  For clarity
let me reiterate:

(1) CPU is 2.8 Ghz with linux kernel 2.4.21  (gcc version 3.2.3)
(2) CPU is 2.4 Ghz with linux kernel 2.6.13  (gcc version 4.0.2)

I started by running the dppc benchmark d.dppc provided by gromacs.  Since
this system does not use PME it is a good test to show performance
without using ffts. The system contains 121856 atoms, and each test run was
5000 steps. The results were:

#procs    time (machine1)      time (machine2)
------    --------------       ---------------
2            1h04:42            1h07:06
6              36:40              35:49
8              27.25              26:33
10             26.26              24:52
12             42.58              19:49
14              --                20.27

The 2 machines appear to behave similarly, and in fact, the newer kernel
appears to make machine (2) perform slightly better despite the slower
CPUs. Machine (1) has optimum performance at 10 CPUs while machine (2)
peaks at 12 CPUs.

I then tested a 51709 atom system (protein + hydrated membrane) using PME.
Each test run was 10000 steps. The results were.

#procs    time (machine1)      time (machine2)
------    --------------       ---------------
2           1h19:19             1h20:07
4           1h01:21             1h37:44
6             56:44             1h59:36
8           1h00:06             2h28:42
10          1h55:21             2h42:31
12          2h47:54             2h56:54

Clearly machine (2) does not scale at all when using PME. I suspected it
was due to fftw3, and compiled using intel mkl, but it doesn't appear to behave
differently. Right now I am at a loss for an explanation, and would
appreciate some help.

Thanks,
David


On Sun, 16 Jul 2006, Mark Abraham wrote:

> >
> > Dear all,
> >
> > I have 2 Intel Xeon clusters:
> >
> > (1) CPU is 2.8 Ghz with linux kernel 2.4.21  (gcc version 3.2.3)
> > (2) CPU is 2.4 Ghz with linux kernel 2.6.13  (gcc version 4.0.2)
> >
> > I run a ~50000 atom system for 500 ps with PME on 8 processors
> > (LAM-MPI) with either machine. With machine (1) the job completes in
> > 12h44:48.  On the other hand, for machine (2) the job does not complete
> > within 24 hours.
> >
> > I am accustomed to such jobs completing in the time demonstrated by
> > machine
> > (1) with the older kernel.
> >
> > Apparently the only major difference is the kernel. Has anyone else
> > experienced anything like this?  Is it possible that fftw works
> > differently
> > under these two systems? Would it be worth it to try intel based ffts?
>
> How about comparing runs that don't use PME, and thus don't use fftw? I
> can't really see the kernel version having any impact... you could compare
> the output from ./configure to see if anything else significant is
> different. The most likely problem is not using the same number of
> processors, but you say you've done that! :-) You could also check whether
> there is a horrible MPI latency problem with machine (2)
>
> Mark
>
> _______________________________________________
> gmx-users mailing list    gmx-users at gromacs.org
> http://www.gromacs.org/mailman/listinfo/gmx-users
> Please don't post (un)subscribe requests to the list. Use the
> www interface or send it to gmx-users-request at gromacs.org.
> Can't post? Read http://www.gromacs.org/mailing_lists/users.php
>
>
> --
>
>