On Thu, 10 Apr 2025 10:02:40 GMT, Thomas Schatzl <tscha...@openjdk.org> wrote:
>> Hi all, >> >> please review this change that implements (currently Draft) JEP: G1: >> Improve Application Throughput with a More Efficient Write-Barrier. >> >> The reason for posting this early is that this is a large change, and the >> JEP process is already taking very long with no end in sight but we would >> like to have this ready by JDK 25. >> >> ### Current situation >> >> With this change, G1 will reduce the post write barrier to much more >> resemble Parallel GC's as described in the JEP. The reason is that G1 lacks >> in throughput compared to Parallel/Serial GC due to larger barrier. >> >> The main reason for the current barrier is how g1 implements concurrent >> refinement: >> * g1 tracks dirtied cards using sets (dirty card queue set - dcqs) of >> buffers (dirty card queues - dcq) containing the location of dirtied cards. >> Refinement threads pick up their contents to re-refine. The barrier needs to >> enqueue card locations. >> * For correctness dirty card updates requires fine-grained synchronization >> between mutator and refinement threads, >> * Finally there is generic code to avoid dirtying cards altogether >> (filters), to avoid executing the synchronization and the enqueuing as much >> as possible. >> >> These tasks require the current barrier to look as follows for an assignment >> `x.a = y` in pseudo code: >> >> >> // Filtering >> if (region(@x.a) == region(y)) goto done; // same region check >> if (y == null) goto done; // null value check >> if (card(@x.a) == young_card) goto done; // write to young gen check >> StoreLoad; // synchronize >> if (card(@x.a) == dirty_card) goto done; >> >> *card(@x.a) = dirty >> >> // Card tracking >> enqueue(card-address(@x.a)) into thread-local-dcq; >> if (thread-local-dcq is not full) goto done; >> >> call runtime to move thread-local-dcq into dcqs >> >> done: >> >> >> Overall this post-write barrier alone is in the range of 40-50 total >> instructions, compared to three or four(!) for parallel and serial gc. >> >> The large size of the inlined barrier not only has a large code footprint, >> but also prevents some compiler optimizations like loop unrolling or >> inlining. >> >> There are several papers showing that this barrier alone can decrease >> throughput by 10-20% >> ([Yang12](https://dl.acm.org/doi/10.1145/2426642.2259004)), which is >> corroborated by some benchmarks (see links). >> >> The main idea for this change is to not use fine-grained synchronization >> between refinement and mutator threads, but coarse grained based on >> atomically switching c... > > Thomas Schatzl has updated the pull request incrementally with two additional > commits since the last revision: > > - * indentation fix > - * remove support for 32 bit x86 in the barrier generation code, following > latest changes from @shade > G1 sets UseCondCardMark to true by default. The conditional card mark > corresponds to the third filter in the write barrier now, and since I decided > to keep all filters for this change, it makes sense to directly use this > mechanism. Do you have performance results for `-UseCondCardMark` vs. `+UseCondCardMark`? The benefit of `+UseCondCardMark` is not obvious from looking at the generated barrier code. ------------- PR Comment: https://git.openjdk.org/jdk/pull/23739#issuecomment-2796872496
