Posted by Valera Veryazov on April 08, 2008 at 22:21:02:
In Reply to: Molcas parallel performance posted by Ignacio on April 08, 2008 at 17:17:55:
Yes. this is known and expected result.
Check the results for alaska, or mckinley - these codes run faster on multicore arch. These codes do not use memory, and I/O.
For all other codes - if you run molcas with several codes - two or more processes fighting for resources - memory and disk.
At the moment molcas parallelization is efficient only for clusters and, to some extend, for SMP.
We started a project to make OpenMP parallelization in Molcas code, but I don't expect that the project will be finished soon - since we need to separate (manually, not via automatic OpenMP parallelization) timeconsuming parts from memory and I/O demanding parts.