AMBER Archive (2008)

Subject: Re: AMBER: AMD Opteron: compiler recommendations?

From: Sasha Buzko (obuzko_at_ucla.edu)
Date: Fri Apr 25 2008 - 17:14:19 CDT

Thanks, Bob.
I'll try to get the Pathscale compilers and check out the difference (if
any).

Sasha

Robert Duke wrote:
> The pathscale compilers are pretty good for opterons; the pgi
> compilers are used for cray machines running opterons, so they must
> not be too bad on performance either, though there has been grief with
> pgi from time to time (in fairness, they have tried to fix their
> problems, so I should give them credit for responding to the past
> problems). With intel, it may be a specific switches problem; I last
> tried running them on opterons a couple of years ago. I preferred
> pathscale for the opteron, but there was not a 30% performance
> differential at that point in time (things could have changed). It is
> completely possible that my default ifort settings for the opteron are
> no longer the best choice (sorry, I just didn't get around to trying
> this combination in the current release cycle).
> Regards - Bob Duke
>
> ----- Original Message -----
> *From:* Sasha Buzko <mailto:obuzko_at_ucla.edu>
> *To:* amber_at_scripps.edu <mailto:amber_at_scripps.edu>
> *Sent:* Friday, April 25, 2008 5:54 PM
> *Subject:* AMBER: AMD Opteron: compiler recommendations?
>
> Hi all,
> I've compiled sander.MPI and pmemd using Intel compilers and
> tested them with no apparent errors on AMD processors. However,
> the executables seem to be considerably faster on Intel chips
> (while the hardware is a bit better in that case, it shouldn't
> account for a 30% performance increase).
>
> Has anyone had any experience with comparing performance of
> binaries built using different compilers on AMD hardware? For
> instance, how do Pathscale compilers compare to Intel on Opterons?
> I've read reports about the evil Intel intentionally
> under-optimizing code on non-Intel chips, but hope it's not the
> issue here :).
>
> Any recommendations and/or benchmark results would be very much
> appreciated.
>
> Thanks in advance
>
> Sasha
>

-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber_at_scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu