On 22 Sep 2010, at 14:03, Gustavo Seabra wrote:
> On Wed, Sep 22, 2010 at 5:02 AM, colvin wrote:
>> Dear all,
>>
>> I am running MD on a system consisting 112550 atoms using pmemd on a
>> quad-core workstation and it tooks about 17 hrs to complete 500ps.
>>
>> Then, i run the same system on 2 quad-core connected with Gigabit
>> ethernet, and it took about 13 hrs.
>>
>> Is this normal? or there are some configurations that can be made for
>> better performance?
>
> So, just to make it clear: your system is running **faster** on the 2
> quad-core/ Gigabit ethernet system than in the 1 quad-core
> workstation, right?
>
> There are a number of competing factors here, and interconnect is just
> one of them. Personally, I don't see anything wrong in your numbers
> (others may correct me): It seems that your system is large enough
> that you are gaining more by spreading it to 8 processors (2
> quad-cores) than what you loose due to the slow interconnect. If you
> really want to test the effect of the slow interconnect, you have to
> restrict the second calculation to use only 2 processes per node for a
> total of 4 processes, as the first calculation.
alternatively time it on the workstation on 1,2,3,4 cores... and then
on the "cluster" (1,2,..,8) and plot the times... if you see a drop
between sys for i<5 cores that tells you something, if the drop is
from i=4 to 5 then that tells you about the interconnect
also, what %age of theoretical peak is this? and how much I/O (or
serial setup) is there (if there's 5 hrs spent doing I/O & setup then
you're comparing 12 to 8 hrs which sounds quite different)
M.
--
Dr. Michael K Bane
RESEARCH COMPUTING SERVICES
Directorate of IT Services
The University of Manchester
www.rcs.manchester.ac.uk
www.twitter.com/mkbane_mcr
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Sep 22 2010 - 06:30:08 PDT
