[gmx-developers] Gromacs benchmark set

[Szilárd Páll]

On Wed, Jul 2, 2014 at 9:28 AM, Alexey Shvetsov <alexxy at omrb.pnpi.spb.ru> wrote:
> Hi Szilárd,
>
> There are actualy few goals related to such benchmarking:
>
> 1. Check if this hardware suitable to run GROMACS (I already did some tests on
> my systems such as RecA protein filaments, nucleosomes etc all have sizes ca
> ~800k - 1.4M atoms). Check scaling and compare it to existing systems (thats
> why I ask about some kind of standart benchmark set)
>
> 2. Another thing is to check scalability of algorithms (PME, RF)
>
> RSC PetaStream is a special systems (it has very similar design to next gen
> Xeon Phi KNL systems). It uses Xeon Phi as regular compute nodes connected
> with multiple InfiniBand links so Xeon Phi to Xeon Phi bandwith is ~6GB/s on
> medium and large size MPI messages.

That's where the catch is, I think. Unless running at a very high
#particles/node (at which the absolute performance would typically
rather low), the typical messages size is <16-32 kB with moderate
parallelization and <4-8 kB below 500 particles/rank.

Some references:
* Fig 2. of [1] shows PP point-to-point comm message sizes with a 134k
system in up to 768 ranks;
* [2] slide 17 shows the message size histogram for dppc bench 121k on
up to 11 nodes; note that on 11 nodes 60% of the messages are <1 Kb
and the rest is in 4-16 kB range too.


I have not seen scaling benchmarks with recent code on Phi clusters,
but knowing how latency-sensitive the code is, I expect that the
unavoidable latency-overhead of PCI-E and especialy PCI-E-IB-PCI-E
should have a strong impact.

> As I see on some systems GMX 5.0 on Xeon Phi scales quite well (up to ~100-200
> atoms per Xeon Phi thread)

How many nodes/cards? With or without PME? With PME how do you place
the PME-only ranks? I expect that putting any on the Phi will
completely kill performance way above 200 atoms/thread - unless you
communicate only locally.

What absolute performance do you get? Can you share the number and perhaps logs?

Cheers,
--
Szilárd

[1] http://www.hector.ac.uk/cse/distributedcse/reports/GROMACS/GROMACS/index.htm
[2] www.hpcadvisorycouncil.com/pdf/GROMACS_Analysis_AMD.pdf

> В письме от 1 июля 2014 17:03:23 пользователь Szilárd Páll написал:
>> Hi Alexey,
>>
>> There is no official benchmark set yet.
>>
>> The right benchmark set will greatly depend on what your goal is.
>> There is a wide range of possible ways to set up benchmarks and there
>> is no single right way to do it. Most importantly, unless the goal is
>> to i) show off hardware ii) benchmark algorithms, the input systems
>> should be representative of the production simulations that are/will
>> be running on the hardware.
>>
>> More concretely, for instance if you want to show decent performance
>> with Xeon Phi (especially strong scaling), you will probably need huge
>> input systems, preferably homogeneous and even better without PME
>> (which  - 3D FFT-s across multiple Phi-s will probably run very
>> poorly). In contrast, if you use an input system like a 50-70k
>> membrane protein simulated with PME, you will probably find it hard to
>> show good performance compared to to an IVB Xeon let alone scaling.
>>
>> IMHO the STFC benchmarks are very disadvantageous for GROMACS (all
>> inputs use CHARMM FF and related peculiar settings) and therefore they
>> are not very representative. Moreover they are outdated too.
>>
>> Cheers,
>> --
>> Szilárd
>>
>>
>> On Mon, Jun 30, 2014 at 3:50 PM, Alexey Shvetsov
>>
>> <alexxy at omrb.pnpi.spb.ru> wrote:
>> > Hi all!
>> >
>> > We're going to run a series of benchmarks on RSC PetaStream system. Its
>> > based on Xeon Phi and designed to run native mode codes. Are there some
>> > kind of representative benchmark set? I'm currently found
>> > http://www.stfc.ac.uk/CSE/randd/cbg/Benchmark/25241.aspx this one. May be
>> > there are some other sets?
>> >
>> > --
>> > Best Regards,
>> > Alexey 'Alexxy' Shvetsov, PhD
>> > Department of Molecular and Radiation Biophysics
>> > FSBI Petersburg Nuclear Physics Institute, NRC Kurchatov Institute,
>> > Leningrad region, Gatchina, Russia
>> > mailto:alexxyum at gmail.com
>> > mailto:alexxy at omrb.pnpi.spb.ru
>> > --
>> > Gromacs Developers mailing list
>> >
>> > * Please search the archive at
>> > http://www.gromacs.org/Support/Mailing_Lists/GMX-developers_List before
>> > posting!
>> >
>> > * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>> >
>> > * For (un)subscribe requests visit
>> > https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-developers or
>> > send a mail to gmx-developers-request at gromacs.org.
>
> --
> Best Regards,
> Alexey 'Alexxy' Shvetsov, PhD
> Department of Molecular and Radiation Biophysics
> FSBI Petersburg Nuclear Physics Institute, NRC Kurchatov Institute,
> Leningrad region, Gatchina, Russia
> mailto:alexxyum at gmail.com
> mailto:alexxy at omrb.pnpi.spb.ru
>
> --
> Gromacs Developers mailing list
>
> * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-developers_List before posting!
>
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>
> * For (un)subscribe requests visit
> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-developers or send a mail to gmx-developers-request at gromacs.org.