Fwd: RFC: TLAB allocation and garbage-first policy

[Roman Kennke]

Does Aleksey's proposed change lead to regressions in other workloads? 
If it improves some workloads, and doesn't regress others, I'm for it.

Can the proposed change be done in a new/other experimental heuristic?

Roman

> On 09/20/2017 06:37 PM, Christine Flood wrote:
>> G1 will trigger a young generation gc when it's designated young regions are full.
> I was talking about mixed collections though, see G1MixedGCLiveThresholdPercent and
> CollectionSetChooser::should_add.
>
>> We trigger a GC when we get to a certain allocation threshold which doesn't really imply anything about how full our current allocation regions are.
> That is not entirely true. Our allocation code walks through freeset linearly, which means that what
> we have left behind as "used" in the freeset includes only fully-allocated and
> racy-fragmented-retired regions...
>
>> Again there is nothing preventing us from writing a heuristic that triggers when we fill up a
>> certain number of regions, but that isn't what we currently do.
> ...so the heuristics that account for ShenandoahAllocationThreshold (which are most of them),
> already do that.
>
> So the difference against G1 is still murky for me. I am experimenting with TLAB trimming.
>
> -Aleksey
>