round up
JuliaLawall opened this issue · comments
It seems that parmap rounds down when statically distributing tasks to cores. For example, I have 284 tasks and 48 cores, which gives 47 cores with 5 tasks each and one core with 49. This represents a pretty massive hit on the benefits of parallelism. Giving 6 tasks to most cores would be a big performance improvement. Everything is fine with ~chunksize:1, but the need for this may not be obvious to most users.
I think chunksize=1 should be the default
everybody wants automatic load balancing to be the default. People who know which chunksize to use should really know what they are doing.
I think most users would benefit from this small little change
I am closing this because there is a workaround. But I agree that the current behavior is sub optimal.
It would have been nice if there was a pull request to implement the suggested change.