[gmx-users] scaling of replica exchange

[Fabio Affinito]

Hi Mark,
I've checked with Valeria and the problem is actually on the setup of
the system (unfair overlapping of the temperature distribution). So I
think that the loss of efficiency should be of a factor between 2 or 3
in her case and not more.

Bye,

Fabio

On 02/23/2011 10:31 AM, Mark Abraham wrote:
> 
> 
> On 02/23/11, *Valeria Losasso * <losasso at sissa.it> wrote:
>>
>> Thank you Mark. I found one message of this month concerning this
>> topic, and there are some small suggestions. I don't think that such a
>> changes can restore a factor of 26, but it could be worth to try to
>> see what happens. I will let you know.
> 
> They won't. The problem is that every 10 (or so) MD steps every
> processor does global communication to check nothing's gone wrong. That
> resulted from some unrelated bits of code trying to share the same
> machinery for efficiency, and treading on each others' toes.
> 
> Mark
> 
>> Valeria
>>
>>
>>
>> On Wed, 23 Feb 2011, Mark Abraham wrote:
>>
>> >
>> >
>> >On 02/23/11, Valeria Losasso <losasso at sissa.it> wrote:
>> >
>> >      Dear all,
>> >      I am making some tests to start using replica exchange
>> molecular dynamics on my system in water. The setup is ok
>> >      (i.e. one replica alone runs correctly), but I am not able to
>> parallelize the REMD. Details follow:
>> >
>> >      - the test is on 8 temperatures, so 8 replicas
>> >      - Gromacs version 4.5.3
>> >      - One replica alone, in 30 minutes with 256 processors, makes
>> 52500 steps. 8 replicas with 256x8 = 2048
>> >      processors, make 300 (!!) steps each = 2400 in total (I arrived
>> to these numbers just to see some update of the
>> >      log file: since I am running on a big cluster, I can not use
>> more than half an hour for tests with less than 512
>> >      processors)
>> >      - I am using mpirun with options -np 256 -s  md_.tpr -multi 8
>> -replex 1000
>> >
>> >
>> >There have been two threads on this topic in the last month or so,
>> please check the archives. The implementation of
>> >multi-simulations scales poorly. The scaling of replica-exchange
>> itself is not great either. I have a working version under
>> >final development that scales much better. Watch this space.
>> >
>> >Mark
>> >


-- 
*********************************************
Fabio Affinito, PhD
CINECA
SuperComputing Applications and Innovation Department - SCAI
Via Magnanelli, 6/3
40033 Casalecchio di Reno (Bologna) ITALY
+39/051/6171794 (Phone)