search

trifork / erjang

A JVM-based Erlang VM
http://www.erjang.org
Apache License 2.0
726 stars 62 forks source link

Improve Scheduler #72

Open krestenkrab opened 10 years ago

krestenkrab commented 10 years ago

Repeatedly, the scheduler turns out to be a bottleneck in performance tests, and -S 1 (schedule everything on one thread) consistently yields better results than -S 2 (or more). I believe that we need a less greedy work-stealing algorithm, so that tasks will tend to stay on the same thread more often; because otherwise we will end up with too many cache misses between the cores.

There must be some simple way to utilize the excellent work by Doug Lea on efficient fork/join and taskstealing.

Mind you, this is really an issue with Kilim (https://github.com/kilim/kilim), who's scheduler we're using.

pron commented 10 years ago

Hi Kresten. You might want to take a look at Quasar. It uses fork/join to schedule its fibers (and the actors running in fibers).

When I have some time I'd like to port erjang to use Quasar rather than Kilim to see the effect on performance. It should be promising, as we're constantly working on improving Quasar.

In any case, I can probably help you if you have any questions regarding FJ. For example, in a ring benchmark, FJ also performs best when using only one thread (and for a good reason).