[Beowulf] re: open64 from AMD (was newbie)

[Rayson Ho]

On Fri, May 8, 2009 at 9:33 AM, Mark Hahn <hahn at mcmaster.ca> wrote:
>> I had no idea that the MIPSPro compiler had been resurrected so I was
>> interested to hear about the AMD effort.

Yeah, PathScale ported Open64 to x86 & AMD64 a few years ago...

And besides IA64 and x64, China (Chinese Academy of Sciences) ported
Open64 to PowerPC & Loongson (the MIPS processor developed by China):

 http://en.wikipedia.org/wiki/Loongson
 http://www.capsl.udel.edu/conferences/open64/2009/


> me too, but it brings up a thorny topic: how to compare the performance
> of different compilers?  the obvious "run it on your code and see" is not
> helpful in my case, since my goal is to have a good choice for a large
> shared HPC organization (several thousand users currently.)

It's hard, and even comparing the performance of the "same" compiler
but different versions can be hard.

Sometimes adding a new optimization can slow down the overall SPEC
mark, and new optimizations with each other, with speedups masking
slowdowns. Making sure that a functionally correct compiler is already
hard enough, and I think it is almost impossible to get a compiler
with no performance regressions. This is especially true when there
are multiple components (different front-ends, optimiziers, backends,
and at the same time supporting multiple processor
micro-architectures), and keep in mind that you (usually) get at least
one new development build each week for each component, and each team
also has a large number of developers checking in code at the same
time!


> how about microbenchmarks?  codes like polyhedron might be reasonable
> to start with.  HPCC?  byte unix benchmarks?

See also: http://developer.amd.com/cpu/open64/AppsAndLibraries/Pages/default.aspx

Rayson




> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
> To change your subscription (digest mode or unsubscribe) visit
> http://www.beowulf.org/mailman/listinfo/beowulf
>