Currently the cluster swapping strategy (in non-deterministic mode) jumps into subclustering as soon as no full cluster swaps are possible on a given rank, irrespective of swapping/non-swapping on other ranks.
This has the effect of improving load balance locally but at the expense of earlier memory locality disruption than as compared to when subclustering is delayed until all ranks have exhausted full cluster swapping possibilities.
This is in contrast to the implementation of temperedlb in vt, which has some global fence between cluster swaps and the subclustering stage.
The goal of this issue is to allow for the latter (vt-like) option in LBAF as well.
Currently the cluster swapping strategy (in non-deterministic mode) jumps into subclustering as soon as no full cluster swaps are possible on a given rank, irrespective of swapping/non-swapping on other ranks.
This has the effect of improving load balance locally but at the expense of earlier memory locality disruption than as compared to when subclustering is delayed until all ranks have exhausted full cluster swapping possibilities.
This is in contrast to the implementation of
temperedlbinvt, which has some global fence between cluster swaps and the subclustering stage.The goal of this issue is to allow for the latter (
vt-like) option in LBAF as well.@lifflander