[gmx-users] GPU efficiency

[Szilárd Páll]

Hi Albert,

I have to concur with Mark's point, it's not hard at all to pick (or
happen to benchmark) circumstances under which one code is
significantly faster than the other.

In order to understand the differences, please share:
- GROMACS logs
[ - full hardware & software specs, compilers, CUDA, etc (will be in
the GROMACS log ;) ]
- full set of AMBER and Desmond setting and algorithms used - there
may be considerable differences in the algorithmic choices that favor
one software or the other.

These details will probably confirm what Mark hinted: a CPU-bound run
will be unfavorable to GROMACS. Depending on the hardware balance
you'll find that mdrun can be slower or faster than AMBER (with
equivalent algorithms). I know because I've compared ;)

On Sun, Jan 24, 2016 at 4:43 PM, Albert <mailmd2011 at gmail.com> wrote:
> Dear Mark:
>
> Thanks for the comments.
>
> I am also very surprised to see the results.  In my impression, Gromacs
> always has the best efficiency for MD GPU running....
>
> I talked this with one of my friends who also uses Gromacs for several
> years. And he told me that he also found similar testing results by himself
> recently....It would be even worse that Gromacs heavily use both CPU+GPU and
> the efficiency is obviously much slower than Amber.

I'm not sure I follow you, but I'm afraid that conclusion could be
incorrect! Could you please clarify what you mean?

One could easily pick a hardware setup that would for instance show
that AMBER-GPU performs worse or some other hardware on which
AMBER-GPU is faster than GROMACS. If you're simply comparing on the
hardware you have (and have made sure to do the correct comparison),
of course, it's going to be the relevant conclusion *for your case*,
but that conclusion does not generalize easily!

> What's need to point out is that Desmond even uses 1 GPU+1 CPU, the speed is
> as fast as Amber running with 2 GPU+2CPU. I contacted Desmond supporters,
> and they told me that they optimize the algorithm recently which fasten the
> running 2 times comparing with previous Desmond release.  As we all known
> that Desmond developer created ANTON which is amazing. It seems that they
> have some breakthrough in GPU these days although currently it only supports
> 1GPU+1CPU for a single job. If in the near future, they somehow manage to
> use multiple GPU for a single job, it will change everything a lot.

Is it multiple time-stepping that they were referring to? AFAIK
Desmond typically uses 2.5 fs time-step and PME every second step. For
a fair comparison you should employ algorithms that accelerate
sampling in the other codes too: virtual sites in GROMACS (up to 2.5x
throughput) and HMR in AMBER (up to 2x throughput). If you did not use
these features, factoring these in will immediately result in much
closer performance numbers.

If there is something else the the Desmond devs came up with, I'd be
curious to hear about it, but the very sparse public information
showed (last time I checked a 1-2 months ago) no performance advantage
wrt. GROMACS.

Cheers,
--
Szilárd

> regards
>
> Albert
>
> On 01/24/2016 09:54 AM, Mark Abraham wrote:
>>
>> Hi,
>>
>> You need to use the hidden -fasterthanamber mdrun benchmarking option!;-)
>> Comparing GROMACS with AMBER on the same hardware is always unfair to one
>> of them, because we use the CPU heavily and they do not. Otherwise we
>> can't
>> comment until we've seen log files shared e.g. on a file sharing service.
>> List won't take attachments.
>>
>> Mark
>
>
> --
> Gromacs Users mailing list
>
> * Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_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-users or send a
> mail to gmx-users-request at gromacs.org.