[TinkersW]

Somewhat interesting read, if rather long..

My experience is that multi-threading has quite abit of overhead, not necessary from the scheduling, but from the cache misses because now everything is unlikely to be in cache, so a naive parallel for can easily end up consuming a ton of CPU resources, it may indeed finish quicker, but use 5x the overall CPU time to do so.

Then there is the other issue that parallel_for suffers from, the "starter" thread has to finish the loop, and if may end up with nothing to do for some time(like when one of the helper threads get suspended..), or it might end up going off to process some other work, causing the entire loop to take much longer to finish. So parallel_for kinda sucks, and I prefer using dependency graphs when I can.