[Pw_forum] abysmal parallel performance of the CP code

[Konstantin Kudin]

 Hi Nicola,

 Looking at how the code scales, it appears that in your case even 8
cpus are absolutely faster than 4, which is not the case for our
cluster. Which MPI library was that?

 Also, I could not test on an empty cluster. However, the timings seem
to be reproducible enough such as the other network traffic seems to
not have too much of an effect in terms of fluctuations from run to
run.

 I have run your AgI test as well, it looks as if something has changed
since the latest code appears to be ~25% slower than it used to be in
Jan '05.

 Kostya


		
__________________________________ 
Yahoo! Mail - PC Magazine Editors' Choice 2005 
http://mail.yahoo.com