Rainer Schuetze
Rainer Schuetze
> My understanding is the evDone is to get all the threads back to sleeping on the evStart event. Yes, it stops the eventual wait(1ms) while the stack is empty...
Hold-off merging, I have run our benchmarks, and this are the results: ``` MIN conalloc 0.752 s, 5 MB, 1965 GC 226 ms, Pauses 223 ms < 0 ms MIN...
> I would think just one event is needed, and all it does is say "there is stuff ready to scan", and it's set when you add something to the...
Switched back to master, as it might not come without side-effects. I also pushed the simplified event handling.
> > Note that an auto-reset event might actually be appropriate in that case! The thread that wakes up can re trigger if there's still data left. > > Indeed,...
> So this is going to be a huge wildcard. Why do we have this loop in there? I guess to get at least some GC activitity. If I disable...
> Maybe up the N value to get more activity? Or limit the number of times this loop can run? > > In any case, I don't think we can...
> I have a bit of concern doing the read of length to be 0 without atomics, and I think the evDone event can just be removed. I think the...
> In a recent DLF meeting, we talked about switching the bump GC to be a full GC interface, so we don't have to hook any of these things, just...
@schveiguy Now pulled out the locking from the scan stack, so evStackFilled could be converted back to being manual reset and busyThreads changed under the same lock. Should be very...