RFR: 8315503: G1: Code root scan causes long GC pauses due to imbalanced iteration [v5]

[Albert Mingkun Yang]

On Tue, 26 Sep 2023 08:42:09 GMT, Thomas Schatzl <tschatzl at openjdk.org> wrote:

>> Hi all,
>> 
>>   please review this change that modifies the code root (remembered) set to use the CHT as internal representation.
>> 
>> This removes lots of locking (inhibiting throughput), provides automatic balancing for the code root scan phase, and (parallel) bulk unregistering of nmethdos during code cache unloading improving performance of various pauses that deal with code root sets.
>> 
>> With a stress test that frequently loads and unloads 6000 classes and associated methods from them we could previously see the following issues:
>> 
>> During collection pauses:
>> 
>> [4179,965s][gc,phases ] GC(273) Evacuate Collection Set: 812,18ms
>> [..]
>> [4179,965s][gc,phases ] GC(273) Code Root Scan (ms): Min: 0,00, Avg: 59,03, Max: 775,12, Diff: 775,12, Sum: 944,44, Workers: 16
>> [...]
>> [4179,965s][gc,phases ] GC(273) Termination (ms): Min: 0,03, Avg: 643,90, Max: 690,96, Diff: 690,93, Sum: 10302,47, Workers: 16
>> 
>> 
>> Code root scan now reduces to ~22ms max on average in this case.
>> 
>> Class unloading (breaking down the code cache flushing, i.e. `CodeCache::flush_unlinked_nmethods`):
>> 
>> Clear Exception Caches   35,5ms
>> Unregister NMethods     598,5ms                                <---- this is nmethod unregistering.
>> Unregister Old NMethods   3,0ms
>> CodeBlob flush           41,1ms
>> CodeCache free         5730,3ms
>> 
>> 
>> With this change, the `unregister nmethods` phase takes ~25ms max on that stress test. @walulyai contributed this part.
>> 
>> We have recently seen some imbalances in code root scan and long Remark pauses (thankfully not to that extreme) in other real-world applications too:
>> 
>> [2466.979s][gc,phases      ] GC(131)     Code Root Scan (ms):           Min:  0.0, Avg:  5.7, Max: 46.4, Diff: 46.4, Sum: 57.0, Workers: 10
>> 
>> 
>> Some random comment:
>> * the mutex for the CHT had to be decreased in priority by one to not conflict with `CodeCache_lock`. This does not seem to be detrimental otherwise. At the same time, I had to move the locks at `nosafepoint-3` to `nosafepoint-4` as well to keep previous ordering. All mutexes with uses of `nosafepoint` as their rank seem to be good now.
>> 
>> Testing: tier1-5
>> 
>> Thanks,
>>   Thomas
>
> Thomas Schatzl has updated the pull request incrementally with one additional commit since the last revision:
> 
>   unregister-nmethod still called during nmethod::flush

If I understand it correctly, this PR consists of two optimizations:

1. Replacing a mutex with conc-hashtable in g1 code-root.
2. Implementing bulk-notify unregistered nmethods.

The first optimization is specific to g1, while the second one touches around the `CollectedHeap::unregister_nmethod` API and could be beneficial for other collectors as well. I wonder if it's possible to split this into two separate PRs for easier reviewing and to gain a better understanding of the performance impact of each optimization.

src/hotspot/share/gc/g1/g1CollectedHeap.cpp line 3041:

> 3039:              p2i(_nm), HR_FORMAT_PARAMS(hr), HR_FORMAT_PARAMS(hr->humongous_start_region()));
> 3040: 
> 3041:       // HeapRegion::add_code_root() avoids adding duplicate entries.

Obsolete comment.

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

PR Review: https://git.openjdk.org/jdk/pull/15811#pullrequestreview-1644078426
PR Review Comment: https://git.openjdk.org/jdk/pull/15811#discussion_r1337092226