RFR: 8310031: Parallel: Implement better work distribution for large object arrays in old gen [v12]

[Albert Mingkun Yang]

On Thu, 28 Sep 2023 07:41:18 GMT, Richard Reingruber <rrich at openjdk.org> wrote:

>> This pr introduces parallel scanning of large object arrays in the old generation containing roots for young collections of Parallel GC. This allows for better distribution of the actual work (following the array references) as opposed to "stealing" from other task queues which can lead to inverse scaling demonstrated by small tests (attached to JDK-8310031) and also observed in gerrit production systems.
>> 
>> The algorithm to share scanning large arrays is supposed to be a straight
>> forward extension of the scheme implemented in
>> `PSCardTable::scavenge_contents_parallel`.
>> 
>> - A worker scans the part of a large array located in its stripe
>> 
>> - Except for the end of the large array reaching into a stripe which is scanned by the thread owning the previous stripe. This is just what the current implementation does: it skips objects crossing into the stripe.
>> 
>> - For this it is necessary that large arrays cover at least 3 stripes (see `PSCardTable::large_obj_arr_min_words`)
>>   
>> The implementation also makes use of the precise card marks for arrays. Only dirty regions are actually scanned.
>> 
>> #### Performance testing
>> 
>> ##### BigArrayInOldGenRR.java
>> 
>> [BigArrayInOldGenRR.java](https://bugs.openjdk.org/secure/attachment/104422/BigArrayInOldGenRR.java) is a micro benchmark that assigns new objects to a large array in a loop. Creating new array elements triggers young collections. In each collection the large array is scanned because of its references to the new elements in the young generation. The benchmark score is the geometric mean of the duration of the last 5 young collections (lower is better).
>> 
>> [BigArrayInOldGenRR.pdf](https://cr.openjdk.org/~rrich/webrevs/8310031/BigArrayInOldGenRR.pdf)([BigArrayInOldGenRR.ods](https://cr.openjdk.org/~rrich/webrevs/8310031/BigArrayInOldGenRR.ods)) presents the benchmark results with 1 to 64 gc threads.
>> 
>> Observations
>> 
>> * JDK22 scales inversely. Adding gc threads prolongues young collections. With 32 threads young collections take ~15x longer than single threaded.
>> 
>> * Fixed JDK22 scales well. Adding gc theads reduces the duration of young collections. With 32 threads young collections are 5x shorter than single threaded.
>> 
>> * With just 1 gc thread there is a regression. Young collections are 1.5x longer with the fix. I assume the reason is that the iteration over the array elements is interrupted at the end of a stripe which makes it less efficient. The prize for parallelization is paid ...
>
> Richard Reingruber has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Remove stripe size adaptations and cache potentially expensive start array queries

Performed additional performance testing on the latest revision of this pull request and https://github.com/openjdk/jdk/compare/master...albertnetymk:jdk:pgc-precise-obj-arr?expand=1 (made sure they were on top of the same master commit)

Couldn't identify significant differences when running micro benchmarks from the JBS ticket with different gc-threads; implemented various tweaks but the distinction between the two approaches remains mostly marginal. Both methods exhibit substantial improvements over the master, as demonstrated earlier.

No performance difference observed in pjbb2005 between the master, this pull request, and shadow-card-table. (I had difficulty in running `timefold`, so I asked Thomas for help about it.)

The cost of malloc + memset for the shadow-card-table is ~0.26ms per 1G of old-gen (each card being 512 bytes) (raw data: 0.553169 ms for 4395946 cards).

Since the shadow-card-table approach doesn't result in any noticeable regression, offers better scalability for large-array-objects, and comes with the lowest implementation complexity, I am inclined to settle for the shadow-card-table approach for now and explore more sophisticated optimizations later on. What are others' thoughts on this direction?

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

PR Comment: https://git.openjdk.org/jdk/pull/14846#issuecomment-1746714371