[gmx-users] OpenMosix and Gromacs

[Erik Lindahl]

Hi,

On Jul 2, 2007, at 11:15 AM, Ansgar Esztermann wrote:

> On Fri, Jun 29, 2007 at 03:49:01PM -0600, Nicolas Sapay wrote:
>>
>> I wonder if someone has some experience with Gromacs and  
>> OpenMosix. Has
>> someone tried recently to run Gromacs in parallel with an OpenMosix
>> cluster? I have read some messages about that on the mailing-list but
>
> I have not tried that; but I do have some experience with OpenMosix
> development, and I would not expect that combination to work well.
> oM copes very well with dynamic loads since it distributed
> processes on the fly, but Gromacs is more likely to generate a static
> load that does not change at all during the simulation run.
> However, some spurious migrations will likely take place even during a
> static load scenario, taking away quite some time (on the order of  
> seconds).
>
> Moreover, oM does not in itself provide any inter-process  
> communication,
> so MPI is needed anyway. There might be a problem with shared memory
> (which is not supported by oM).
>
> Finally, most syscalls are very slow under oM since they are  
> executed on
> the home node. This pertains especially to file operations, which go
> through the network unbuffered.
>
>
> All in all, oM was written to solve a problem that usually does not
> occur when running programs like Gromacs, so there seem to be many
> potential problems but little benefit from combining those two.
>

At least in the "old days", Mosix did not support migration sockets,  
i.e. processes doing communication. Not sure if that has been solved,  
but since Gromacs just uses MPI calls for communication we never have  
any knowledge about node names, IP numbers etc - that's entirely up  
to the MPI library.

Thus, if an MPI implementation works fine on Mosix Gromacs will too,  
and the CVS 3.99 branch will load balance beautifully. If MPI doesn't  
work there's nothing we can do.

Cheers,

Erik