RFR: 8354727: CompilationPolicy creates too many compiler threads when code cache space is scarce [v3]
Manuel Hässig
mhaessig at openjdk.org
Thu Jun 19 14:41:43 UTC 2025
On Wed, 18 Jun 2025 18:27:17 GMT, Manuel Hässig <mhaessig at openjdk.org> wrote:
>> Running
>>
>>
>> java -XX:+SegmentedCodeCache -XX:ReservedCodeCacheSize=10M -XX:NonNMethodCodeHeapSize=6M \
>> -XX:ProfiledCodeHeapSize=5M -XX:NonProfiledCodeHeapSize=5M -version
>>
>>
>> on a machine with more than 285 cores, this would fail with the message that the specified `NonNMethodCodeHeapSize` is too small to fit all compiler buffers (instead of failing because the sum of the heaps is larger than the `ReservedCodeCacheSize`). Hence, the calculated compiler count is too high. This is due to CompilationPolicy::initialize() checking how many compiler buffers fit into the `ReservedCodeCacheSize`. However, in the case above, this is significantly larger than `NonNMethodCodeHeapSize` (especially on a debug build) and causes a check changed in #17244 to fail. That check was changed to check that all compiler buffers fit into the `NonNMethodCodeHeap` instead of the `NonNMethodCodeHeap` having at least a size of `CodeCacheMinimumUseSpace`.
>>
>> # Changes
>>
>> This PR fixes the calculation of the `CICompilerCount` ergonomic. Firstly, @shipilev kindly provided a fix for the compiler buffer size used in the calculation is also correct if we only have C2. Secondly, `NonNMethodHeapSize` is used as the maximum buffer size available for compilers buffers in the calculation of the maximum number of compiler threads instead of `ReservedCodeCacheSize`. Therefore, the check failing in the explanation above can never fail because we set the number of compiler threads only so high that they will always fit into the `NonNMethodCodeHeap`.
>>
>> This change changes how many compiler threads are created by the `CICompilerCount` ergonomic. For the default value `NonNMethodCodeHeapSize=5M`this limit is 24 compiler threads on a system 285 cores or more for product builds and 20 threads for debug builds on a system with 145 cores or more.
>>
>> # Testing
>>
>> - [x] [Github Actions](https://github.com/mhaessig/jdk/actions/runs/15733154809)
>> - [x] tier1 through tier3 plus Oracle internal testing on our supported platforms
>
> Manuel Hässig has updated the pull request with a new target base due to a merge or a rebase. The pull request now contains 14 commits:
>
> - Merge branch 'master' into JDK-8354727-policy
> - Calculate buffer size correctly for c2_only
>
> Co-authored-by: Aleksey Shipilev <shade at openjdk.org>
> - Caclulate how many compiler buffers fit into NonNMethodCodeHeap
> - Clarify endif
> - update copyrights
> - remove leftover <type_traits> include
> - fix whitebox access to code cache size configs
> - VMPageSizeConstraintFunc
> - CodeCacheMinBlockLength
> - CodeCacheExpansionSize
> - ... and 4 more: https://git.openjdk.org/jdk/compare/7bc0d824...a05f9bd3
> This undid cf78925 changes. Something wrong was with merge.
I fixed that in the latest commit. Thank you for pointing it out.
-------------
PR Comment: https://git.openjdk.org/jdk/pull/25872#issuecomment-2988331468
More information about the hotspot-compiler-dev
mailing list