[CP2K:638] Re: cp2k speedup on multicore machines

Fawzi Mohamed fa... at gmx.ch
Fri Jan 25 16:15:39 UTC 2008


I know that I have an almost perfect scaling with 32 H2O going from 1  
to 2 cpus.

You have to check what you are measuring (be sure to measure real time).

You have to be sure that the mpi correctly sets cpu affinity (so that  
the process do not swap processors and get slow).
I know lammpi and openmpi, both do it, I assume mpich also does it.

With lammpi you have to write the host filelike this
myhost  cpu=2

and not
myhost
myhost

ciao
Fawzi
On Jan 25, 2008, at 3:28 PM, Teodoro Laino wrote:

> The problem, unfortunately, is far being so obvious..
>
> I think the MB is just mistyping.. (GB instead..)
> And the problem is not only related to the benchmark input but also  
> to small systems.
> I suggested  Luigi to move to the benchmark test since in that case  
> (a larger one)  you should definitely see scaling (more difficult
> to observe with smaller systems). IO is reduced to the minimum in  
> testing.. so don't consider the access to the disk.
>
> Luigi, I just post the info you gave me about compilers/hardware..  
> I don't have the compilation flags.. Can you post them here in the  
> group?
>
> ---------------------------------------------------------------------- 
> ---------------------------------------------------------------------- 
> -----
> Linux k119 2.6.21-1.3194.fc7 #1 SMP Wed May 23 22:47:07 EDT 2007  
> x86_64 x86_64 x86_64 GNU/Linux
> gcc version 4.1.2 20070925 (Red Hat 4.1.2-27)
> ifort Build 20070613 Package ID: l_fc_c_10.0.025
> intel mkl 10.0.1.014
> fftw-3.1.2
> mpich2-1.0.6p1
> ---------------------------------------------------------------------- 
> ---------------------------------------------------------------------- 
> -----
>
> We use, on our local machines mostly the same setup.. The only  
> difference is the kernel  (ours is a 2.6.16.21-0.8-smp, could this  
> be an issue Axel?) and the mpich2 (we use 1.0.5p4)...
>
> Definitely it's not an OMP issue.. following Manuel suggestion I  
> asked Luigi to set OMP_NUM_THREADS=1 and nothing changes in the  
> timings..
>
> Let's see if we can figure out what's going on..
> Sure.. it's a little bit weird...
> Teo
>
> On 25 Jan 2008, at 15:06, Fawzi Mohamed wrote:
>
>> well 8MB RAM is for sure a bottleneck ;)
>> 8GB should not, but, well 6 Hours for 10 steps... maybe it did  
>> swap, and if you swap then everything is over, do check, if memory  
>> consumption is more than your physical memory the cpu usage drops  
>> drastically, again in top you should see this.
>>
>> Use a smaller system, if this is the problem, the 32 water should  
>> already scale well...
>>
>
>
> >

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20080125/f2dfc79a/attachment.htm>


More information about the CP2K-user mailing list