[Beowulf] Re: Beowulf Digest, Vol 45, Issue 28

[Geoff Jacobs]

Håkon Bugge wrote:
> At 21:00 17.11.2007, Chris Samuel <csamuel at vpac.org> wrote:
>> I should also point out that we're using the PGI 7.x series of
>> compilers and tell our users to build with:
>>
>> -tp k8-64,barcelona-64
>>
>> so that they get the optimisations for both dual and quad core chips
>> (with the appropriate path selected at run time).
> 
> While I did the SPEC MPI 2007 for the Barcelona, I got the following
> projections (8 nodes, 8 cores per node) for SPECmpiM_base2007:
>    * PathScale 3.0, -march=core: 5.53
>    * PathScale 3.0, -march=opteron -msse3: 5.32
>    * PGI 7.0-7, -tp barcelona-64 + lots of fancy switches: 4.71
>    * PGI 7.0-7, -tp barcelona-64 -nomovnt -fast: 4.75
> I didn't have much time to play with compiler options. It was
> interesting to see that the "wrong" architecture was better using the
> PathScale compiler. Further, PathScale was significant better than PGI
> 7.0-7, even though PGI had support for the "right" architecture.
> 
> The switches used for the PathScale compiler can be found at the SPEC
> web, http://www.spec.org/mpi2007/results/res2007q4/
> 
> The final rate (5.61) was better, due improved memory bandwidth (due to
> a better BIOS).

No big secret, I don't think. Core optimized code will be scheduled for
a wider SIMD unit, same as Barcelona.

-- 
Geoffrey D. Jacobs

To have no errors
  would be life without meaning
  No struggle, no joy