Skip to content

WIP: Repartition using CPU ticks

Peter W. Draper requested to merge repart-by-ticks into master

Change repartitioning so that we only use CPU ticks as the weights.

In tests this seems to be sufficient and can give the best quality partitions, so reinstate the use of ticks to weight the graph that is partitioned by METIS.

The changes are to reintroduce measuring CPU ticks for all tasks by default (we keep --enable-task-debugging as that requires other fields), but removing any attempt to force repartitioning, we only do it when enough information about the tasks is available, i.e. after steps that interact most particles.

This decouples task weights from repartitioning as scheduling needs to be done with partial information, for that we still have an issue with tasks that do not have characteristic functions that scale with particle counts.

Edited by Peter W. Draper

Merge request reports