I'm hoping someone can help shed some light on an issue that's
cropped up during some benchmarking. I'm not looking for any comments
along the lines of 'that's why Opteron' are better' etc. I'm just
trying to find a solution to a problem that's come up.
First, the application is pure FORTRAN 77. It has been compiled with
either XLF on a dual G5 1.8 GHz with 2 GB RAM under Panther or with the
64-bit Portland compilers on a dual Opteron 240 (1.43 GHz), 8GB RAM
with 64-bit Linux (mandrake, I believe). During no time in the
execution of the program does paging occur (as observed from top at
least).
Here is the issue. It appears that scalar portions of the code
specifically FFT calculations on the Opteron system are significantly
faster than on the G5. When the code enters a region where there is
branching, however, the two systems normalize and the G5 slightly
outperforms the Opteron.
In looking through the Portland manual, it states that on a 64-bit OS
with Opterons the floating point instructions default to SSE/SSE2. I
interpret this to mean that it's using the FPU's on the vector units.
Presumably this means that there are more registers there for
processing more data (correct me if I'm wrong). This is the only thing
that I could ascribe the very fast floating point operations to on the
Opteron.
Is my interpretation of what is occurring correct, or is the
performance difference due to something entirely different? If it is
the case, are there compiler directives or procedures that can be used
to increase the floating point performance (throughput?) on the G5 via
Altivec? That is, without going through and hand vectorizing all of
the various routines that are slow.
One other thing that I'll point out that caused me to think it was
the SSE/SSE2 usage. If I use the FFT's in vDSP for the portion of the
FFT calculation are 2x faster than on the Opteron. If I compile the
application in 32-bit mode on the Opteron the G5 FFT ends up being 4-6x
times faster.
I'd appreciate any insight that anyone could offer or solutions that
might help boost performance on our G5's.
Thanks in advance,
Dave
David W. Gohara, Ph.D.
Harvard Medical School
http://www.scianafilms.com
617-432-1216 (p)
617-432-4360 (f)
