Re: efficiency of parallelization?

Posted by Jesper Wisborg Krogh on October 14, 2002 at 11:01:36:

In Reply to: efficiency of parallelization? posted by Darko Babic on October 13, 2002 at 19:16:29:

In MOLCAS version 5.2 there isn't much gained by running a parallel calculation, because only SEWARD, SCF, and ALASKA have been parallized. On the other hand, we will release version 5.4 in a few weeks (will be free for 5.0 or 5.2 license holders). Here MCKINLEY and MCLR will pe parallized as well.

Our own experience when it comes to efficiency it that you should gain the following speed ups:

SCF: Super-linear for semi direct calculations and close to linear for direct calculations.

MCKINLEY: Very close to linear.

MCLR: A bit worse than McKinley.

The speed ups of course requires the calculation to be relatively big compared to the number of nodes, and the benchmark tests of MCKINLEY and MCLR are still not finished.

Follow Ups:

Post a Followup

Name:
E-Mail:

Subject:

Comments:
: In MOLCAS version 5.2 there isn't much gained by running a parallel calculation, because only SEWARD, SCF, and ALASKA have been parallized. On the other hand, we will release version 5.4 in a few weeks (will be free for 5.0 or 5.2 license holders). Here MCKINLEY and MCLR will pe parallized as well. : Our own experience when it comes to efficiency it that you should gain the following speed ups: : SCF: Super-linear for semi direct calculations and close to linear for direct calculations. : MCKINLEY: Very close to linear. : MCLR: A bit worse than McKinley. : The speed ups of course requires the calculation to be relatively big compared to the number of nodes, and the benchmark tests of MCKINLEY and MCLR are still not finished.