[gmx-users] Pme on gpu performance

Szilárd Páll szilard.pall at cbr.su.se
Tue Nov 29 21:49:24 CET 2011 

Hi,

On Mon, Nov 28, 2011 at 8:55 PM, Jones de Andrade <johannesrs at gmail.com> wrote:
> Hi.
>
> I'm another one waiting for the OpenCL version here.
>
>> 4.6 will support only CUDA.
>>
>> OpenCL GPU implementations have yet match CUDA; on CPU it would hardly
>> match the performance of the hand-tuned SSE kernels, moreover, CPU &
>> GPU algorithms are not the same.
>
> If you allow me, the (second and third!) argument is not strong enough,
> since also cuda for cpu would not match the performance of all tricks from
> gromacs. Also, I think that the cpu and cuda algorithms are not the same.
> From my point of view, the only really strong argument would be the first
> one: "OpenCL GPU implementations have yet match CUDA!", despite the fact
> that I would like to see it better explained if possible (even a link would
> be suffice).

I'll try give clear and complete replies on why no OpenCL, but no
flaming, please. ;)

1) OpenCL has only recently gotten reasonably stable and complete both
the a standard as well as implementations from the various vendors. At
the time when we started this work, OpenCL implementations could match
CUDA in stability/features, let alone performance. This has changed
since, but only fairly recently; e.g. NVIDIA OpenCL 1.1 support was
only introduced AFAIR with driver v280.09-13 or so.

Additionally, clusters with AMD GPUs are rare as a hen's teeth and
even if they will appear overnight, porting to OpenCL is nowhere near
as difficult as coming up with fast and future-proof algorithms.

So I hope it's clear now that by "match" I didn't only mean
performance, but also stability, HPC adoption, support, etc.


2/3) I don't really see the point you were trying to make by comparing
CUDA->X86 to the CPU performance of Gromacs.

IMHO CUDA for X86 is not of very much use unless performance is
irrelevant (e.g. emulation). If you want fast CPU code, you need to
write the algorithm & performance-aware implementation for CPUs, a
CUDA->X86 compiler won't do the hard lifting. E.g. coalesced access
for GPUs is pretty much the worst access pattern on CPUs as it ensure
nice and high cache miss rate.
The same applies to OpenCL: you need *different* kernels for CPUs and
GPUs - or more precisely, one kernel per vendor per hardware
generation. Moreover, I have yet to see a case study where OpenCL on
CPU matches the performance of optimized SSE code (I hope somebody
will surprise me soon).

No doubt that there will be OpenCL (GPU) support in Gromacs sometime
in the near future, but not in 4.6. However, if there are many users
out there with floating-point beast AMD cards, let us know. What comes
after 4.6 is not set in stone at all.

Cheers,
--
Szilárd


> More important: does this mean that openMM and gromacs gpu development are
> "breaking apart"? What can we really expect from that, other than easier
> installation and usage?
>
> Thanks a lot!
>
> Sincerely yours,
>
> Jones
>
>
>>
>> --
>> Szilárd
>>
>>
>>
>> On Sun, Nov 27, 2011 at 11:25 PM, Alexey Shvetsov
>> <alexxy at omrb.pnpi.spb.ru> wrote:
>> > Hi!
>> >
>> > Will it use CUDA or OpenCL? Second one will be more common since it will
>> > work with wider range of platfroms (cpu, gpu, fcpga)
>> >
>> > Szilárd Páll писал 27.11.2011 23:50:
>> >>
>> >> Native acceleration = not relying on external libraries. ;)
>> >>
>> >> --
>> >> Szilárd
>> >>
>> >>
>> >>
>> >> On Sun, Nov 27, 2011 at 9:33 PM, Peter C. Lai <pcl at uab.edu> wrote:
>> >>>
>> >>> Random hijack:
>> >>>
>> >>> Will the CUDA acceleration support CMAP or will it still be limited to
>> >>> the limitations of OpenMM that it currently has?
>> >>>
>> >>> On 2011-11-27 12:10:47PM -0600, Szilárd Páll wrote:
>> >>>>
>> >>>> Hi Andrzej,
>> >>>>
>> >>>> GROMACS 4.6 is work in progress, it will have native CUDA
>> >>>> acceleration
>> >>>> with multi-GPU support along a few other improvements. You can expect
>> >>>> a speedup in the ballpark of 3x. We will soon have the code available
>> >>>> for testing.
>> >>>>
>> >>>> I'm a little skeptical about the 5x of ACEMD. What setting did you
>> >>>> use? AFAIR they were using large time-steps without contraints in
>> >>>> their benchmarks.
>> >>>>
>> >>>> Cheers,
>> >>>> --
>> >>>> Szilárd
>> >>>>
>> >>>>
>> >>>>
>> >>>> On Thu, Nov 24, 2011 at 9:53 AM, Andrzej Rzepiela
>> >>>> <Andrzej.Rzepiela at physik.uni-freiburg.de> wrote:
>> >>>> > Dears,
>> >>>> >
>> >>>> > I would like to ask what performance improvement is expected for
>> >>>> > future
>> >>>> > releases (e.g. 4.6) for PME calculations on gpu cards. Last week I
>> >>>> > sent
>> >>>> > benchmark results for dhfr to the list, which show no performance
>> >>>> > gain
>> >>>> > when
>> >>>> > the newest versions of tesla cards vs intel xeon R are used (more
>> >>>> > or
>> >>>> > less in
>> >>>> > agreement of what is presented on the gromacs website). However,
>> >>>> > for
>> >>>> > the
>> >>>> > same system  simulations with different gpu code, acemd,  on a very
>> >>>> > similar
>> >>>> > machine can be about 5 times faster, as benchmarked recently by my
>> >>>> > colleagues, suggesting that performance of gromacs may soon
>> >>>> > improve.
>> >>>> > is that
>> >>>> > correct ?
>> >>>> >
>> >>>> >
>> >>>> > Andrzej
>> >>>> > --
>> >>>> > gmx-users mailing list    gmx-users at gromacs.org
>> >>>> > http://lists.gromacs.org/mailman/listinfo/gmx-users
>> >>>> > Please search the archive at
>> >>>> > http://www.gromacs.org/Support/Mailing_Lists/Search before posting!
>> >>>> > Please don't post (un)subscribe requests to the list. Use thewww
>> >>>> > interface
>> >>>> > or send it to gmx-users-request at gromacs.org.
>> >>>> > Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>> >>>> >
>> >>>> --
>> >>>> gmx-users mailing list    gmx-users at gromacs.org
>> >>>> http://lists.gromacs.org/mailman/listinfo/gmx-users
>> >>>> Please search the archive at
>> >>>> http://www.gromacs.org/Support/Mailing_Lists/Search before posting!
>> >>>> Please don't post (un)subscribe requests to the list. Use the
>> >>>> www interface or send it to gmx-users-request at gromacs.org.
>> >>>> Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>> >>>
>> >>> --
>> >>> ==================================================================
>> >>> Peter C. Lai                    | University of Alabama-Birmingham
>> >>> Programmer/Analyst              | KAUL 752A
>> >>> Genetics, Div. of Research      | 705 South 20th Street
>> >>> pcl at uab.edu                     | Birmingham AL 35294-4461
>> >>> (205) 690-0808                        |
>> >>> ==================================================================
>> >>>
>> >>> --
>> >>> gmx-users mailing list    gmx-users at gromacs.org
>> >>> http://lists.gromacs.org/mailman/listinfo/gmx-users
>> >>> Please search the archive at
>> >>> http://www.gromacs.org/Support/Mailing_Lists/Search before posting!
>> >>> Please don't post (un)subscribe requests to the list. Use the
>> >>> www interface or send it to gmx-users-request at gromacs.org.
>> >>> Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>> >>>
>> >
>> > --
>> > Best Regards,
>> > Alexey 'Alexxy' Shvetsov
>> > Petersburg Nuclear Physics Institute, Russia
>> > Department of Molecular and Radiation Biophysics
>> > Gentoo Team Ru
>> > Gentoo Linux Dev
>> > mailto:alexxyum at gmail.com
>> > mailto:alexxy at gentoo.org
>> > mailto:alexxy at omrb.pnpi.spb.ru
>> > --
>> > gmx-users mailing list    gmx-users at gromacs.org
>> > http://lists.gromacs.org/mailman/listinfo/gmx-users
>> > Please search the archive at
>> > http://www.gromacs.org/Support/Mailing_Lists/Search before posting!
>> > Please don't post (un)subscribe requests to the list. Use the www
>> > interface
>> > or send it to gmx-users-request at gromacs.org.
>> > Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>> >
>> --
>> gmx-users mailing list    gmx-users at gromacs.org
>> http://lists.gromacs.org/mailman/listinfo/gmx-users
>> Please search the archive at
>> http://www.gromacs.org/Support/Mailing_Lists/Search before posting!
>> Please don't post (un)subscribe requests to the list. Use the
>> www interface or send it to gmx-users-request at gromacs.org.
>> Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>
>
> --
> gmx-users mailing list    gmx-users at gromacs.org
> http://lists.gromacs.org/mailman/listinfo/gmx-users
> Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/Search before posting!
> Please don't post (un)subscribe requests to the list. Use the
> www interface or send it to gmx-users-request at gromacs.org.
> Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>

More information about the gromacs.org_gmx-users mailing list