RFR: 8252104: parallel heap inspection for ShenandoahHeap [v3]

Lin Zang lzang at openjdk.java.net
Fri Sep 18 09:43:29 UTC 2020


On Fri, 18 Sep 2020 08:03:13 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:

>> Lin Zang has refreshed the contents of this pull request, and previous commits have been removed. The incremental views
>> will show differences compared to the previous content of the PR.
>
> src/hotspot/share/gc/shenandoah/shenandoahHeap.cpp line 1470:
> 
>> 1468:     for (uint j = 0; j < roots_num; j++) {
>> 1469:       uint stack_id = j % _num_workers;
>> 1470:       oop obj = _roots_stack.pop();
> 
> I am not sure we can `pop()` the stack from multiple threads without any sort of synchronization. It might be better to
> just walk the same roots from multiple threads, and then let the bitmap help us figure out if we have visited the root
> already. See how `ShenandoahVerifierReachableTask` does it?

Hi @shipilev,
Thanks for your effort for reviewing this change!
The method prepare_worker_queue() here is invoked in the constructor of ShenandoahParallelObjectIterator, which is
designed to run in single thread.  The logic of using ParallelObjectIterator is as follows: (refer to code in
hotspot/share/memory/heapInspection.cpp)
 -  create instance of ParallelObjectIterator in shenandoahHeap:
      ParallelObjectIterator* poi = Universe::heap()->parallel_object_iterator(gang->active_workers())
 - create heap iteration task and run (invoke ParallelObjectIterator's method object_iterator())
     ParHeapInspectTask task(**poi**, cit, filter);
     // Run task with the active workers.
     gang->run_task(&task);
=======
So the ShenandoahParallelObjectIterator iterate all roots in the single thread, collecting all roots in the
roots_stack, then prepare the queues for every workers.  -- processed in constructor. And the workers start their
processing from their queue, and then iterate all reachable objects. -- processed in work(uint worker_id)

I have checked that ShenandoahVerifierReachableTask supports iterating roots parallelly, I didn't choose that way
because I want to make every workers to start with nearly the same number of roots in their queue, hope it could help
on inital work load balance.

Lin

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

PR: https://git.openjdk.java.net/jdk/pull/67


More information about the serviceability-dev mailing list