[Pw_forum] speedup of md simulation

HW Sheng hwsheng at jhu.edu
Fri Dec 13 17:06:52 CET 2002


Dear Paolo,

Thanks for the reply.

I understand that our network communication is a drag. We use 100M fast
Ethernet. As to the planewave parallelization, I have tried other ab intio
software package like CPMD. It seems the speedup there is much greater
(around a factor of 5 on my 10 dual processor cluster). does the different
performance stem from different methodologies that are utilized in those
simulation packages?

By the way, the new pwscf web page looks much nicer.

Thanks.

Howard

> [...] I installed pwscf on my Linux/alpha cluster (10 dual
processors),which
> employs LAM/MPI [...] I found that the speedup is only a factor of 2
> [...]
> (2) is this due to the fact that the code is not optimally parallelized
>     or that the LAM/MPI is not well configured?

...or none of the above? unless your cluster has high-speed communication
hardware (i.e. Myrinet, not Ethernet), you will not get a good speedup.
You may even get "speedown" instead.

>  [...] PS: When I do k-point scf calculation, the speedup is acceptable,
> around a factor of 5 or so.

you mean parallelization on k points ? it requires much less communication
than parallelization on plane waves. See also
http://www.pwscf.org/performance.htm#parallelization






More information about the users mailing list