[gmx-users] Re: MPICH on Gbit with > 4 CPUs?

[Justin MacCallum]

> has anybody ever managed to run gromacs in parallel on more than 4  
> CPUs
> when using MPICH on Gigabit Ethernet? I have just searched the mailing
> lists, where this problem was also reported (but no solution!).
>
> We have dual-CPU Xeons, Linux 2.4. With up to 4 CPUs, gromacs 3.3 and
> 3.2.1 work. With 5+ CPUs, mdrun initializes properly but never  
> finishes the first
> time step. An MPI test program (that does an all-to-all communication)
> works happily on up to 40 CPUs with MPICH. With LAM, mdrun works fine
> on any number of CPUs as well. Why not with MPICH?

Hi Carsten,

I have experienced the same problem you mention. I haven't been able  
to resolve it yet. I've noticed some very strange behavior regarding  
the distribution of processes. Some of the nodes seem to have too  
many mdrun process while other nodes only have one (when they should  
have two). Also, some of the nodes have 100% CPU usage, while others  
are sitting completely idle. Unfortunately its been quite a while  
since I tested this so I don't exactly remember the specifics.

--------------------------
Justin MacCallum
PhD Student
Departement of Biological Sciences
University of Calgary

email: jlmaccal at ucalgary.ca
web: moose.bio.ucalgary.ca