It strongly depends on the particular calculation and on the parallelization strategy you choose. Linear scaling would be the ideal scaling which you never get. The things to keep in mind are load balancing and communication overhead. This last is affected by your communication network latency and bandwidth ... As a general rule the bigger the calculation the easier is to scale. It may be useful to have a look at the final timing summary of a calculation to see the fraction of time spent in communication (FFT scatter/gather and reduce operations) compared with the total wall clock time for your system and how this changes for different settings. Another factor to keep in mind is RAM memory as with certain parallelization strategies you can trade some speed with increased memory.
stefano mohnish pandey wrote: > Dear QE users, > I am trying to see how the runtime scales with > number of processors. I did the same calculation using one node with eight > processors on one cluster and 8 nodes with total 56 processors on other > cluster but the time does not seem to scale linearly with the number of > processors. Can anybody give me an idea how does the time scale with number > of processors. > Thanks a lot in advance. > > > ------------------------------------------------------------------------ > > _______________________________________________ > Pw_forum mailing list > Pw_forum at pwscf.org > http://www.democritos.it/mailman/listinfo/pw_forum >
