[gmx-users] cpu utilization

[Erik Lindahl]

Hi Dean,

Performance depends on many things, in particular the type of system 
you are simulating and whether all bonds or only the ones to hydrogens 
are constrained. Your case sounds like load balancing issues.

Version 4.0 should see significantly improved automatic load balancing 
over many nodes even for small systems; Gromacs scales for large 
systems (try the DPPC benchmark, for instance), but in prior version 
we've mainly been focused on pushing single-CPU performance as far as 
possible. Since you need statistics, we could almost always use several 
independent runs to utlize clusters to the max.

Cheers,

Erik

>
> Also, perhaps a bit of evidence, it has a real hard time cleaning up 
> after itself. The output indicating performance and stuff comes *way* 
> before the jobs eventually die and some never die. I had to write a 
> simple  'killgro' command to nuke everything for good. Any ideas?
>
Sounds like a problem with your MPI implementation or file systems - 
once Gromacs writes the last stuff we essentially just do an 
MPI_Finalize() call.

Cheers,

Erik