Nick McNaughton

Cosmos Solve Scalability

Discussion created by Nick McNaughton on Jul 1, 2008
We've been looking into workstation configurations for solving CosmosWorks studies as quickly as possible. For the most part it seems that clock speed is king, especially when using the direct sparse solver. However, I've got a fairly specific question that I'd appreciate someone with access to a quad core machine having a go at.

We're seeing 75-80% CPU usage while converging on a solution with the iterative solver on a dual core machine, and that leaves us wondering whether more cores would appreciably reduce solution time for that phase of the solve. If someone with a quad core machine could do a back-to-back solve with 4 cores and 2 cores (set cpu affinity for star.exe in task manager) noting solution time and approximate CPU usage for each, that'd be really interesting information. If you're keen, a 3 core solve would complete the set nicely, too.. and anyone with dual quad xeons is welcome to step up to the plate.

For a trial problem, I'd suggest some sort of contact problem with ~200k nodes - say a steel plate with a hole in it, and a pin in the hole with no penetration contact conditions. That should iterate long enough to get some decent solve time results.