Closed qiemem closed 5 years ago
The initial evidence for this came from running LS models on NUs Quest cluster. We found that running with more virtual processors allocated to NetLogo actually resulted in slower runs. However, In retrospect, I believe the models we were running didn't actually leverage LS's parallelism anyway. Thus, this needs to be re-evaluated.
I've confirmed that this problem was actually due to the way we were distributing NetLogo on Quest nodes and actually has nothing to do with LevelSpace. I've since fixed the job dispatch script. BehaviorSpace actually synchronizes pretty regularly anyway, so until we see evidence of this being a problem, I'm going to close this.
Since BehaviorSpace already does much better parallelization than LS can, we should just run single threaded in BehaviorSpace. Especially in a cluster, since communication between processes becomes more expensive, running with more threads can actually make things slower. This is a super bad problem.