Re: [AMBER] Recommendations for Amber on AMD Bulldozer

From: Jan-Philip Gehrcke <jgehrcke.googlemail.com>
Date: Tue, 21 Aug 2012 19:17:42 +0200

Ross,

thanks for your comments.

Just for completion of this thread: the bulldozer platform is quite
special. The OS's task scheduler needs to be adjusted to Bulldozer's
core/module layout in order to maximize performance. For Windows, this
had been addressed via a hotfix. I am not entirely sure, but it looks
like as if all the Bulldozer optimizations for Linux have been applied
only to >=3.1 and not to the 2.x branch.

Regarding some new instructions, it looks like only GCC >=4.7 and Open64
>=4.5 have implemented them so far (and make use of it when
-march=bdver1 and an optimization level >0 is specified).

It would be cool to have a benchmark comparing Amber 12 on Bulldozer
under different conditions (Linux 2.x vs. 3.2, Intel vs. GCC vs Open64,
different optimization levels). Maybe at some point we will be able to
come up with such a study.

Still, it would be great to hear about experiences others have with
Amber on Bulldozer.

All the best,

Jan-Philip








On 08/21/2012 04:56 PM, Ross Walker wrote:
> Hi Jan-Philip,
>
> My advice would just be to use the Intel compilers. These will likely give
> you the best performance and save you the hassle of trying to configure
> pgi etc or try to use some other weird untested compiler.
>
> I wouldn't bother with ACML either - just stick with MKL it should work
> fine. ACML was mostly single precision last time I looked so not much use
> anyway.
>
> All the best
> Ross
>
>
> On 8/21/12 5:02 AM, "Jan-Philip Gehrcke" <jgehrcke.googlemail.com> wrote:
>
>> Hello Amber community,
>>
>> In Dresden we would like to build Amber 12 for running on an AMD
>> Bulldozer cluster (Opteron 6274 "Interlagos").
>>
>> In theory, it looks like the following setup would be good:
>> - Open64 compiler with -march=bdver1 and -O3 (cf.
>> http://developer.amd.com/Assets/CompilerOptQuickRef-62004200.pdf)
>> - ACML
>>
>> On the other hand, we could just go for the GNU or PGI way.
>>
>> I see that the configure script of Amber 12 now supports open64, but
>> that this is still in an 'experimental' state. In the archives of this
>> list I read some pretty old discussions about using Amber with ACML and
>> my impression was that it's not worth the effort.
>>
>> Do you have any recommendations or experiences regarding compiler,
>> compiler flags, and math library for the Bulldozer platform? We have to
>> trade off build simplicity against performance of the outcome. Regarding
>> Amber 12, does it make sense to build against ACML? Would it be stupid
>> to not use Open64 on this platform?
>>
>> Thanks for your insights,
>>
>> Jan-Philip
>>
>>
>>
>>
>>
>> _______________________________________________
>> AMBER mailing list
>> AMBER.ambermd.org
>> http://lists.ambermd.org/mailman/listinfo/amber
>
>
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>


_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Aug 21 2012 - 10:30:05 PDT
Custom Search