[gmx-users] Re: workload gromacs on dual processor

[Erik Lindahl]

>
>
>>
>
>I have exactly the same problem, when running mdrun on a (single node) dual
>Xeon processor Linux machine. The total CPU usage (40% + 60% or 50% + 50% on
>CPU 0 and CPU1) adds up to 100% of a single processor. This doesn't seem to
>be as efficient as it should be, although I think it's still very fast
>compared to other methods!
>You say that you have to prepare the input for multiple processors (in this
>case 2), but how can I do this?: the grompp and mdrun have a -np option,
>which prepares the input for multiple nodes, but I do not find an option for
>peparing the input for multiple processors on a single node.
>

Hi,

First - the CPU usage on a dual machine will depend on the other jobs 
running on it - by default Gromacs uses nice 19,
so if anything else is running it will yield unless you specify a 
difference nicelevel.

There isn't really any difference between multiple CPU and multiple node 
jobs right now; the parallelization uses multiple jobs (one per CPU) 
that communicate with MPI calls. It doesn't matter if you run them all 
on one node or on different nodes, so you should prepare the files for 
two nodes!

(The reason we changed it to "nodes" is to prepare for the future 
split-level approach)

Cheers,

Erik