AMBER12 QM/MM: is d-orbital code going to be parallelized?

his message: [ Message body ] [ More
options (top, bottom) ]
Related messages: [ Next message ] [ Previous
message ] [ In reply to ] [ Next
in thread ] [ Replies ]

Contemporary messages sorted: [ by date ] [ by
thread ] [ by subject ] [ by author ]
[ by messages with attachments ]

From: filip fratev <filipfratev.yahoo.com>

Date: Fri, 20 Apr 2012 20:35:25 -0700 (PDT)

Hi,

I made an additional test with larger (more realistic) QM region.

The QM region consist of 164 atoms (ATP,Mg2+ and several Myosin residues);6-31g/rhf; 3 minimization steps:

TeraChem 1GPU=231.65sec; 2GPU's=129.1sec (1.8x, but will be around 2x after 100 steps)

Gaussian 4
cores .4.6Ghz= 1321.2sec;

Thus we have 10.2x gain in
performance for that system and the Amber-TeraChem link works well. I run only 3 steps and if we run 100 the speed up will be likely 15x. I did mistake in first
test and now the results between TeraChem and Gaussian are very similar.

Hope that this
will help to those that
interest in QM/MM simulations with Amber
and TeraChem!

All the best,

Filip

________________________________

From: filip fratev <filipfratev.yahoo.com>

To: AMBER Mailing List <amber.ambermd.org>

Cc: Andreas Goetz <agoetz.sdsc.edu>

Sent: Friday, April 20, 2012 8:30 AM

Subject: Re: [AMBER] AMBER12 QM/MM: is d-orbital code going to be parallelized?

Dear Prof. Case
and Andreas,

Thank you for your responses!

I use the last TeraChem version - 1.5.

According to Amber 12 manual there is no limitations, as such for GAMMES, for

electronic embedding in TeraChem. Are you sure that with this version it is

possible only mechanical embedding or I should I ask Petachem developers about

that?

�

I performed some tests using TeraChem and

Gaussian as external programs to assess mainly the GPU speed up by TeraChem. I

used two GTX580 3GB cards and 2600K.4.6Ghz CPU. Some results and questions:

�

1) Indeed all Amber tests passed.

2) Secondly I tested the example from Tutorial

A2. Both programs work well with Amber link and gave very similar results using

various parameters. In this test (only 12 QM atoms) Gaussian was faster than

TeraChem (TC) for minimization. In the case of MD simulation I observed only

20% boost in the performance using TC. Moreover, TC was faster with one GPU

instead of two (30% slower with 2 GPU's). This is likely due to the small

system size. In fact using Gaussian I was not able to achieve some good

CPU-core scale too.

3) Further I performed some more realistic tests

using Myosin protein (MM) + ATP-Mg2+ (QM). The system is around 116K atoms

solvated. Here TeraChem had a clear advantage but not exactly what I expected. During

the 20 steps minimization TC was 4.75x faster than Gaussian (all cores) and

during MD 4.3x. It is indeed impressive because for only 20 steps MD (it was

6-31G, but I do not remember b3lyp or something else) it took 705sec using Gaussian

and only 163sec by TeraChem; 4,9ps vs 21.2ps/day. Here again the GPU scale is

the main problem - the second GPU adds only 20% speed up. If I use only

TeraChem for optimization it is 100%, thus one can expect around 8 x speeds up.

What could be the problem? The software link or

something else? ��

��

I tested (not only now) the GPU speed up when

use only TeraChem for optimization or just SP and it is exactly 2x for two

GPU's, doesn't matter for 30 or 150 atoms. What are your advices and in general

is it possible to be improved GPU speed up for the Amber-TeraChem combination?

�

I am not expert in QM/MM calculations that's

why would like to ask you what semi-empirical method in Amber will best fit for

ATP-Mg2+ system and what is the best way, in your opinion, to describe the

energy difference of �ATP-Mg2+ binding if

mutations are present in the ATP binding pocket?

�

All the best,

Filip

________________________________

From: Andreas Goetz <agoetz.sdsc.edu>

To: AMBER Mailing List <amber.ambermd.org>

Cc: filip fratev <filipfratev.yahoo.com>

Sent: Thursday, April 19, 2012 10:50 PM

Subject: Re: [AMBER] AMBER12 QM/MM: is d-orbital code going to be parallelized?

On Apr 16, 2012, at 1:08 PM, David A Case wrote:

> On Mon, Apr 16, 2012, filip fratev wrote:

>

>> I'd like to extend the question. According to

>> the Amber 12 manual it is possible to perform QM/MM using TeraChem software

>> (Cuda software). However, developers of the TeraChem ensured me that this not

>> possible at the moment and probably will be possible in June.

>

> What happens when you try such calculations?� In particular, do the TeraChem

> test calculations work?� (See, e.g.� amber12/test/qmmm_EXTERN/QMMM_MD_TeraChem).

>

>> Thus, which programs are available at the

>> moment for QM/MM calculations?

>

> Please see Section 3.7 of the Amber12 Reference Manual.

>

> ....dac

For electronic embedding you need a TeraChem version that supports external point charge electric fields. You can try mechanical embedding if this fits your needs. This should work also with earlier versions of TeraChem. Check if following test works:

$AMBERHOME/test/qmmm_EXTERN/QMMM_MD_TeraChem/Run.nma-spcfw-15.hf_sto-3g.MechEm

Be aware that I have not extensively tested mechanical embedding.

If you are using the interface to external QM programs, any feedback is appreciated.

All the best,

Andy

--
Dr. Andreas W. Goetz
Assistant Project Scientist
San Diego Supercomputer Center
Tel� : +1-858-822-4771
Email: agoetz.sdsc.edu
Web� : www.awgoetz.de
_______________________________________________
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