[gmx-users] parallel simulation in dual core PC

[Mark Abraham]

On 21/04/2010 11:29 PM, Hsin-Lin wrote:
> Hi,
>
> I compare the run times.
> The calculation of two processors is faster than one.
> And I saw the utility of the two cores large than 85% separately.
> I can sure it's exactly parallel simulation in my PC now.
>
> Another question is:
> someone told me the interconnect, 1G Ethernet, of the machine in our lab. is
> too slow to do parallel simulation.
> What the effect should be in this kind of slow Ethernet environment?

GROMACS will run very slowly, because it needs low latency 
communication. Gigabit ethernet is too slow here. See papers by Carsten 
Kutzner, for example.

Mark

> Is it why the utility of CPU are always less than 35% when I submit the same
> job to the machine in our lab.?
>
> Hsin-Lin
>> Hi,
>>
>> I want to use my dual core PC to do the simulation.
>> The version of GROMACS installed is version 4.0.5.
>> According the message post by others on mailing list before.
>> I type "lamboot" first, and type the commend:
>> mpirun -np 2 mdrun_mpi -s 200ns.tpr -e 200ns.edr -o 200ns.trr -g 200ns.log
>> -c 200ns.gro.
>> When the simulation starts, I use the commend, "top", to check the utility
>> of CPU.
>> I found there is only one core which is used by GROMACS.
>> And the utility of this core is separated to 52% and 48% for two mdrun_mpi
>> jobs.
>
> Depending what is being measured by what, this could be a correct
> result. We can't really tell because we don't know how you've set up
> LAM, or how your top works.
>
> Do a short test calculation with one and two processors and compare the
> run times.
>
> Mark
>