RFR: 8319048: Monitor deflation unlink phase prolongs time to safepoint [v7]

[Erik Österlund]

On Thu, 16 Nov 2023 20:07:51 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:

>> See the symptoms, reproducer, and analysis in the bug.
>> 
>> There is a major problem in current unlinking code: we only check for safepoint every 1M monitors, which might take a while with large population of dead monitors, prolonging time to safepoint. Even if we spend 1ns per monitor, that's already +1ms in TTSP. In reality, we see double-digit-ms outliers in TTSP. (There is a secondary problem that comes with searching for new `prev` if monitor insertion happened while we were preparing the batch for unlinking; this might theoretically take the unbounded time.)
>> 
>> This PR fixes the issue by providing a smaller batch size for unlinking. The unlinking batch size basically defines two things: a) how often do we check for safepoint (`ObjectSynchronizer::chk_for_block_req` in the method below); and b) how much overhead we have on mutating the monitor lists. If we unlink monitors one by one, then in a worst case, we would do a CAS on `head` and the atomic store for `OM._next` for every monitor, both of which are expensive if done per monitor.
>> 
>> The experiments with the reproducer from the bug shows that the threshold of 500 works well: it mitigates TTSP outliers nearly completely, while still providing the large enough batch size to absorb list mutation overheads. See how bad outliers are in baseline, and how outliers get lower with lower batch, and almost completely disappear at 500. I believe the difference between baseline and `MUB=1M` is short-cut-ting the search for new `prev`.
>> 
>> ![plot-monitor-unlink-1](https://github.com/openjdk/jdk/assets/1858943/21b22029-6802-44a6-aea5-52b7fd124717)
>> 
>> Additional testing:
>>  - [x] Linux AArch64 server fastdebug, `tier1 tier2 tier3`
>>  - [x] Ad-hoc performance tests, see above
>
> Aleksey Shipilev has updated the pull request with a new target base due to a merge or a rebase. The incremental webrev excludes the unrelated changes brought in by the merge/rebase. The pull request contains 11 additional commits since the last revision:
> 
>  - Merge branch 'master' into JDK-8319048-monitor-deflation-unlink
>  - Merge branch 'master' into JDK-8319048-monitor-deflation-unlink
>  - Move unlink_batch init to proper place
>  - Add invariant checking
>  - Merge branch 'master' into JDK-8319048-monitor-deflation-unlink
>  - Fix test after recent test infra renames
>  - Merge branch 'master' into JDK-8319048-monitor-deflation-unlink
>  - Batch 500
>  - Pre-final touchups
>  - Option range and tests
>  - ... and 1 more: https://git.openjdk.org/jdk/compare/014e1899...09863092

So today we:
1. Walk monitors from the in-use list, try to deflate monitors, and build an external data structure of the monitors we deflated
2. Walk said deflated monitors again through said external data structure, and unlink them from the in-use list
3. Handshake
4. Walk said deflated monitors again through said external data structure, and unlink them from the in-use list

The external data structure can't deal with more than max_jint elements, which forces us to do some kind of batching not to blow up the data structure.

I would propose to add another _purge_next field to the ObjectMonitor, so that step 1 can immediately unlink from the in-use list, as well as add to the purge list, in one go, making step 2 that we are trying to batch, completely redundant. That seems to remove the need for the new JVM flag as well as the existing MonitorDeflationMax that really was just added to deal with GrowableArray length being an int. Step 1 already does incremental polling, so then if we just add incremental polling to step 4 just as we did in step 1, we could release weak handles in the ObjectMonitor destructor, and call it a day. How does that sound? Essentially what I propose is:

1. Walk the in-use list with incremental polling, if we deflate a monitor, unlink it from in-use list and add to purge list
2. Handshake
3. Walk the purge list with incremental polling, delete ObjectMonitor (with destructor that releases its WeakHandle)

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

PR Comment: https://git.openjdk.org/jdk/pull/16412#issuecomment-1816503951