24 Sep
2014
24 Sep
'14
10:18 a.m.
Dear Tuomas, a somewhat late reply. On 11/08/14 22:49, Tuomas Karna wrote: ...
Load balancing may be an issue on this grid, haven't looked at that yet. I'm also seeing performance degradation over time in parallel, so I suspect there's a bug somewhere but haven't dug deep enough to find it.
Eventually, we found this bug, there was a problem in the way we were preallocating matrices in parallel when running non-linear problems that had zero initial conditions, resulting in performance degradation when assembling Jacobians. If you get a chance to run with current firedrake [352295c] and PyOP2 [adb4114] master branches, you should hopefully see that this has gone away. Cheers, Lawrence