search

invenia / Parallelism.jl

A library for threaded and distributed parallelism.
MIT License
8 stars 1 forks source link

reduced overhead with a static schedule? #4

Open ChrisRackauckas opened 4 years ago

ChrisRackauckas commented 4 years ago

https://github.com/invenia/Parallelism.jl/blob/master/src/tmap.jl#L10

That would have quite a bit of overhead. It would be nice to have a static schedule (@threads) version for fast running programs, which would then be nice once there's an adjoint definition for the map.

oxinabox commented 4 years ago

This is largely solved by using the basesize kwarg. https://github.com/invenia/Parallelism.jl/blob/159a138d562a1583e44129340a0dfa784a34c523/src/tmap.jl#L36

It won't get to be fully as fast as using @threads might, but it would get relatively close.

In general current implementation of tmap assumes the work being done takes tens of seconds, or even minutes. Thus even when the time taken to do each piece of work is theoretically identical, various other external factors are likely to put it out of sync anyway, and so one wants to use a basesize that doesn't fully partition the work