Hi Jan-Philip,
> you've shown a significant change of "double GPU" performance in
> dependence of CPU performance.
>
> I am wondering if this dependency can also exist for "single GPU"
> simulations. How can I find out if my CPU is limiting the performance
> of
> a GPU simulation without trying different CPUs? Is the CPU usage an
> indicator? I've been running a simulation of a small system on a single
> C2070 and saw the pmemd process creating a CPU usage of 100 %. Do these
> 100 % already suggest that the CPU limits the performance?
The way pmemd.cuda is written is such that for serial GPU calculations the
entire calculation is run on the GPU, the CPU does nothing except sit
spinning at barriers waiting to do I/O. As such the CPU speed and memory
bandwidth (and to some extent the speed of the PCI-E slot) will have little
to no impact on the GPU code performance for a single GPU run.
The CPU is at 100% utilization because it is basically sitting there
spinning waiting for interrupts hence this doesn't mean anything with
regards to performance.
In parallel (using more than one GPU for a single run) the situation is very
different. Here the GPUs have to communicate on every time step. Hence the
PCI-E connection speed needs to be at least x16 PER GPU and the higher the
memory bandwidth of the CPU memory (related to the front side bus speed of
the CPU) the better.
All the best
Ross
/\
\/
|\oss Walker
---------------------------------------------------------
| Assistant Research Professor |
| San Diego Supercomputer Center |
| Adjunct Assistant Professor |
| Dept. of Chemistry and Biochemistry |
| University of California San Diego |
| NVIDIA Fellow |
|
http://www.rosswalker.co.uk |
http://www.wmd-lab.org/ |
| Tel: +1 858 822 0854 | EMail:- ross.rosswalker.co.uk |
---------------------------------------------------------
Note: Electronic Mail is not secure, has no guarantee of delivery, may not
be read every day, and should not be used for urgent or sensitive issues.
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Jan 11 2012 - 07:30:02 PST
