RFR: 8319650: Improve heap dump performance with class metadata caching [v3]

[Aleksey Shipilev]

On Thu, 9 Nov 2023 10:45:18 GMT, Aleksey Shipilev <shade at openjdk.org> wrote:

>> See the rationale and some profiles in the bug.
>> 
>> This PR implements simple caching for object instance dumping, which is overwhelmingly hottest path in heap dumping code. This allows to put the cache straight in `HeapObjectDumper`, which is effectively thread local. A simple hash table would then work to cache the class metadata.
>> 
>> Given a simple workload from the bug, these are the motivational improvements and historical perspective:
>> 
>> 
>> $ for I in `seq 1 3`; do java -XX:+UseParallelGC -XX:+HeapDumpAfterFullGC -Xms8g -Xmx8g HeapDump.java 2>&1 | grep created; rm *.hprof; done
>> 
>> # Baseline JDK 8u382
>> Heap dump file created [1440865228 bytes in 5.492 secs]
>> Heap dump file created [1440865228 bytes in 5.460 secs]
>> Heap dump file created [1440865228 bytes in 5.471 secs]
>> 
>> # Baseline JDK 11.0.21
>> Heap dump file created [1446043070 bytes in 5.088 secs]
>> Heap dump file created [1446042039 bytes in 5.096 secs]
>> Heap dump file created [1446042835 bytes in 5.081 secs]
>> 
>> # Baseline JDK 17.0.10
>> Heap dump file created [1446841531 bytes in 3.276 secs]
>> Heap dump file created [1446841673 bytes in 3.342 secs]
>> Heap dump file created [1446841590 bytes in 3.323 secs]
>> 
>> # Baseline JDK 21.0.2 EA
>> Heap dump file created [1447828697 bytes in 5.516 secs]
>> Heap dump file created [1447829756 bytes in 5.461 secs]
>> Heap dump file created [1447829539 bytes in 5.435 secs]
>> 
>> # Baseline JDK 22
>> Heap dump file created [1447285149 bytes in 5.232 secs]
>> Heap dump file created [1447285149 bytes in 5.211 secs]
>> Heap dump file created [1447284813 bytes in 5.222 secs]
>> 
>> # Patched JDK 22
>> Heap dump file created [1447285087 bytes in 1.071 secs]
>> Heap dump file created [1447285149 bytes in 1.073 secs]
>> Heap dump file created [1447285073 bytes in 1.058 secs]
>> 
>> 
>> This amounts to 5x improvement, 1.2 GB/sec dump rate (somewhat I/O limited), and the best heap dump performance across all JDK releases. I am planning to backport this patch where possible: at least to JDK 21u, but hopefully to JDK 17u as well.
>> 
>> There are other simple improvements, but this one is the top bottleneck.
>
> 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 three additional commits since the last revision:
> 
>  - Limit the table size
>  - Merge branch 'master' into JDK-8319650-heapdump-cache-meta
>  - Fix

A bit more realistic test, `spring` from Dacapo 23.11:


$  java -XX:+UseParallelGC -XX:+HeapDumpAfterFullGC -Xms256m -Xmx256m -jar dacapo-23.11-chopin.jar spring

# Before
Dumping heap to java_pid61605.hprof ...
Heap dump file created [21075711 bytes in 0.060 secs]
Dumping heap to java_pid61605.hprof.1 ...
Heap dump file created [35894740 bytes in 0.116 secs]
Dumping heap to java_pid61605.hprof.2 ...
Heap dump file created [65576913 bytes in 0.518 secs]
Dumping heap to java_pid61605.hprof.3 ...
Heap dump file created [74486004 bytes in 0.547 secs]
===== DaCapo 23.11-chopin spring starting =====
Starting 8192 requests...
Dumping heap to java_pid61605.hprof.4 ...
Heap dump file created [85493155 bytes in 0.597 secs]
Completing query batches: 100%
Completed requests
===== DaCapo 23.11-chopin spring PASSED in 5009 msec =====

# After
Dumping heap to java_pid61545.hprof ...
Heap dump file created [21009027 bytes in 0.023 secs]
Dumping heap to java_pid61545.hprof.1 ...
Heap dump file created [35898640 bytes in 0.041 secs]
Dumping heap to java_pid61545.hprof.2 ...
Heap dump file created [65457100 bytes in 0.072 secs]
Dumping heap to java_pid61545.hprof.3 ...
Heap dump file created [74483982 bytes in 0.089 secs]
===== DaCapo 23.11-chopin spring starting =====
Starting 8192 requests...
Dumping heap to java_pid61545.hprof.4 ...
Heap dump file created [85815887 bytes in 0.115 secs]
Completing query batches: 100%
Completed requests
===== DaCapo 23.11-chopin spring PASSED in 4539 msec =====

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

PR Comment: https://git.openjdk.org/jdk/pull/16545#issuecomment-1803626788