[Beowulf] Thought that this might be of interest

[Joe Landman]

Jeffrey B. Layton wrote:

[...]

> So which one is faster - Woodcrest or Opteron - really
> depends on your code.

Yup.

Most of the testing I have done, where Woodcrest is faster appears to be 
linked to the cache size.  Since I used the PGI compilers, and built a 
single binary (not the mixed binary), it sort of pulls this out and 
highlights it.

A CPU with 4M apparent cache for a single processor will run circles 
around a machine with a 1M apparent cache per processor for cache 
friendly apps running a single thread.  As you increase the number of 
threads per socket, things even out.  And this jives well with my 
testing to date.

That said, I am quite interested in trying to feed the Woodcrest SSE 
pipeline at a better rate than the compilers allow for (same with the 
Opteron).  Unfortunately this gets into the area of hand coded SSE*, and 
the Intel compiler does a decent job supporting this (while the PGI 
doesn't).  Since the Intel compiler has other things working against it 
(ahem!), this makes the job of writing fast code somewhat harder.

I know, lets all just use GCC ...



-- 

Joseph Landman, Ph.D
Founder and CEO
Scalable Informatics LLC,
email: landman at scalableinformatics.com
web  : http://www.scalableinformatics.com
phone: +1 734 786 8423
fax  : +1 734 786 8452 or +1 866 888 3112
cell : +1 734 612 4615