thanks!

[Steve]

Thanks to everyone on the list for their help -- after lots of
reading, a replaced DIMM, and a couple of replaced symlinks, my SC072
is happily crunching away on some magnetohydrodynamics codes. Next: N-
body (particle-particle) and smooth-particle-hydrodynamics (SPH).

First: does anyone have info or case studies on parallel timings? I
was playing around with some simple MPI codes yesterday, and found
good scaling, but am interested to know more about tuning codes for
MIPS processors generally and the SC072 in particular.

Cheers, and happy holidays,
Steve

[Philip Mucci]

Hi Steve,

If you have a copy of the most recent programmers guide, there is
quite a good chapter on the performance tool infrastructure available
for the machine.

To get you started, you can probably do:

mpirun <args> papiex -a <your app> and examine the summary files
produced...

Tuning for this particular processor is largely oriented around the
memory pipeline, particularly cache reuse. The prefetch instruction is
somewhat anemic as well, so you may have some luck disabling that.
Also, it's worth experimenting between -O2, -O3, -O3 -OPT:Ofast and -
Ofast... And don't forget -ffast-math at the lower levels, but be sure
to check your results. Also, it goes without saying, the pathscale
compiler is better than GCC at just about everything, barring heavily
templated C++ code. One last thing, if you've ported this code, you
should make sure you're not taking floating point exceptions via
denorms or other FP badness... They're hard to spot and the symptoms
can easily be confused with TLB traps. So if you're not seeing 100% in
user cycles, you may want to look those those... If you're not
familiar with denorms or other causes of FP traps, a simple google
search should be helpful.

Feel free to contact me with other questions...

Happy holidays.

Philip Mucci
Business Development
phil....@samara-llc.com
+1 (415) 577-9899

> --
>
> You received this message because you are subscribed to the Google
> Groups "SiCortex Users" group.
> To post to this group, send email to sicorte...@googlegroups.com.
> To unsubscribe from this group, send email to sicortex-user...@googlegroups.com
> .
> For more options, visit this group at http://groups.google.com/group/sicortex-users?hl=en
> .
>
>