[gmx-users] Re: Re: Re: Question about parallazing Gromacs

[Cesar Araujo]

> ------------------------------
>
> Message: 7
> Date: Thu, 14 Sep 2006 09:27:45 +0200
> From: "Qiao Baofu" <qiaobf at gmail.com>
> Subject: Re: [gmx-users] Re: Question about parallazing Gromacs (Qiao
> Baofu)
> To: "Discussion list for GROMACS users" <gmx-users at gromacs.org>
> Message-ID:
> <6a91f07b0609140027u68ac3085tb01120f19b854545 at mail.gmail.com>
> Content-Type: text/plain; charset="iso-8859-1"
>
> Hi,
>
> Thanks. I have test different cpus. Our institute has two clusters: one is
> each node has 4 cpu (A), one is one node has only 1 cpu (B).   I made
> different tests on the two clusters and my local computer using the same
> system. See the following result:
>
> A     (For 1 hour)
>    # of cpus   ;    MD steps
>      4             finished (200000steps for 26:21)
>      8             finished (200000steps for 40:57)
>     12             87950
>     20             42749
>     44             5962   !!!!!
> B   (For 1 hour)
>     # of cpu           ; MD steps
>       1                156991  for 56:12
>       2                179820
>       3                200,000 for 54:20
>       4                200,000 for 51:12
> c. Local(single cpu), 200000 steps  For 1h52:38
>
> One can see that
> 1. On cluster A, one nodes(4 cpu) is just as 4 times fast as my local
> computer.
> 2. More than one nodes will decrease the performancs the gromacs,
> 3. On cluster B, the more cpu used, the faster gromacs runs. But the
> difference of speed is not apparent.
> 4. Cluster B with 4 cpus is slow as half as that Cluster A with 1 node (4
> cpus)
>
> I wonder if anyone can tell the bottlenack: the hardware on the cluster or
> gromacs?
>
>

Well, probably it is a combination of bottlenecks in both software and 
hardware.
Both of them should be optimized to guarantee the best performance. From the
hardware side the delays in inter-node communication across the network can
be the reason. But also, it can be related with software issues related to 
the operating
system and network configuration. I think that the first thing to do is work 
in cluster
setup to achieve the best performance for your hardware configuration.

In addition, take into account that inter-node communication between two 
CPU's
inside the same machine should be faster than between two CPU's across the 
network.

>From the software side (Gromacs) the algorithms should be optimized with 
regard
to the parallel programming model. David said in a previous answer that they
have detected some problems in the software implementation and they are 
working
to solve that issue in the next version of Gromacs.

Regards,
-----------------------------------------------------------
Cesar Araujo, Lic. of Chemistry
Department of Molecular Endocrynology
Oulu University Hospital
FIN-90029 OYS, OULU, FINLAND

phone: +358 8 3155632
e-mail: cesar.araujo at oulu.fi