[gmx-users] Reg: Running gromacs on a condor cluster

[Mark Abraham]

On 3/05/2010 7:55 AM, Florian Dommert wrote:
>
> On 30.04.2010, at 16:04, Esztermann, Ansgar wrote:
>
>> On Apr 30, 2010, at 14:46 , Rohit Farmer wrote:
>>>
>>> Hi everyone..
>>>
>>> I just made a small two node condor cluster and was trying to run gromacs on it ... so i used the vanilla environment and placed the gromacs mdrun command in a shell script and used the shell script to run the command ... and it is running also ... but only one one system .. my pool is showing four processor on doing condor_status ... but only one processor is clamed ... not all ... so if anyone works on gromacs and condor both then please help ...
>>
>> Does it work locally, i.e. without Condor? After all, Condor is just a batch system. It will help you start mdrun on one or more nodes of your cluster, but it does not magically parallelize things.
>> You would either use an MPI distribution or pthreads. (I am not sure if the latter is already available in a release version.)
>> How do you call mdrun? Do you use mpiexec? What are the parameters?
>>
>>
>
> Another issue is the configuration of Condor. Usually the vanilla universe is not capable of running parallel jobs. However as Ansgar mentioned, if only one process starts it is most probable that you did not use the command mpiexec responsible to start the requested number of processes on the system.
>
> So if all works on your machine, just use the corresponding command line in a bash script and submit it to a Condor universe that is capable to treat parallel jobs, requesting the appropriate number of processes with machine_count.

... and even if it works, you still have to show a profit from the 
parallelism. Even gigabit ethernet is normally not useful enough, even 
without any Condor overhead.

Mark