RFR: 8302215: G1: Last-ditch Full GC should do serial compaction for tail regions in per thread compaction points. [v2]

[Albert Mingkun Yang]

On Wed, 15 Feb 2023 15:44:19 GMT, Ivan Walulya <iwalulya at openjdk.org> wrote:

>> Fwiw, I tend to prefer the `sort` call as it is least amount of code that achieves the goal well enough. All these other suggestions seem to be optimizations for little gain that need lots of extra code/helper data structures for something not (time) critical.
>> 
>> (I may not completely understand @albertnetymk 's suggestion: without either sorting the regions after all of them have been filtered from the worker's compaction point which involves the same sorting, or doing the sorted insertion of the regions for each worker thread's cp regions which is unlikely to be faster than the overall (quick)`sort` at the end, the suggestion does not seem to work).
>
> I also prefer the sort. Common case, the number of regions in serial == ~no of threads.

> Fwiw, I tend to prefer the sort call as it is least amount of code that achieves the goal well enough. All these other suggestions seem to be optimizations for little gain that need lots of extra code/helper data structures for something not (time) critical.

I think what I suggested requires less code than the sorting approach.

> Common case, the number of regions in serial == ~no of threads

True for now, but when humongous regions can be moved, this number can go up.

-------------

PR: https://git.openjdk.org/jdk/pull/12529