Closed evanberkowitz closed 8 years ago
This is done. The current behavior is this:
if no possible tasks were successfully launched on one pass over the remaining work:
update the MAX_{NODES,GPUS} = min( 2*MAX_{NODES,GPUS}, {NODES,GPUS} ) respectively
update the MIN_{NODES,GPUS}= MIN_{NODES,GPUS} / 2
loop again
In some circumstances the user may want to do many small tasks rather than a few large tasks. Currently there is no way for the user to express / impose this desire.