RFR: Fix wb-assist for Traversal

[Roman Kennke]

Am 24. Oktober 2018 08:17:36 GMT-07:00 schrieb Aleksey Shipilev <shade at redhat.com>:
>On 10/24/2018 05:11 PM, Roman Kennke wrote:
>> WB-assist is currently not working for Traversal. It is watching out
>for
>> marked objects close to the one just evacuated, but that wouldn't
>work
>> for traversal GC because it's only ever marking evacuated object,
>> nothing in cset gets marked.
>> 
>> In order to fix this, we need to actually do a little sub-traversal
>> starting from the object just evacuated. That's what the patch does:
>> 
>> http://cr.openjdk.java.net/~rkennke/traversal-wb-assist/webrev.00/
>
>I am against adding more complicated code that handles traversal
>assists. This is one of the reasons
>why there is no concurrent-mark assist as well. Not only this code is
>complicated, it also has
>interaction with gc-worker traversal.
>
>Unless this has clear performance benefits, I'd back off doing this.

OK. I'll run more benchmarks when I'm back home. The motivation for this was the realization how heavy the wb slowpath is when I fixed the false-sharing in oom-evac-handler, and then found that the current implementation of wb-assist is currently not doing anything useful. If this patch doesn't show benefits we must at least disable wb-assist for traversal so to not hurt ourselves.

Roman

>
>-Aleksey

-- 
Diese Nachricht wurde von meinem Android-Gerät mit K-9 Mail gesendet.