Make Recycler faster on OpenJ9 #13357
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
franz1981
approved these changes
Apr 22, 2023
|
Previously we were getting these numbers from With this PR we're getting these numbers: So OpenJ9 improves a lot, but OpenJDK regresses. |
Motivation: To avoid reference processing overhead while at the same time ensuring that we clear Recycler data structures that were used by now-terminated threads, we need to check if the LocalPool owner thread has terminated. We used to do this by calling `Thread.getState`, but this turns out to be quite expensive on OpenJ9, because they need to halt the target thread before its state can be safely inspected. See the `haltThreadForInspection` call in [`Java_java_lang_Thread_getStateImpl`](https://github.com/eclipse-openj9/openj9/blob/55b9267740559406b8471000cd3a3b6f681dc095/runtime/jcl/common/thread.cpp#L76-L102). This turns out to be a scalability bottleneck in some deployments. Modification: We can get the same semantics as `Thread.getState() == Thread.State.TERMINATED` by checking `!Thread.isAlive()`. The `isAlive` call is cheap on both [OpenJ9](https://github.com/eclipse-openj9/openj9/blob/03967d8cc24dc1f60168110c9d207d2dff57765a/jcl/src/java.base/share/classes/java/lang/Thread.java#L756-L761) and [OpenJDK](https://github.com/openjdk/jdk/blob/master/src/java.base/share/classes/java/lang/Thread.java#L1798-L1815). Result: We no longer see the scalability bottleneck observed in netty#13347 (comment)
chrisvest
force-pushed
the
41-recycler-opt
branch
from
96454c0
to
3498d7a
Compare
|
We can get the best of both by checking if we're on OpenJ9: |
I don't see any other calls to |
|
I am surprised that we have regressed on Jdk, it doesn't seem anything but a load there - you checked on x86? |
|
I only checked on aarch |
What hardware? Such a heavy impact for a volatile read is odd. |
|
@theRealAph M1 Pro, macOS |
|
@chrisvest @theRealAph https://github.com/AdoptOpenJDK/openjdk-jdk8u/blob/9a751dc19fae78ce58fb0eb176522070c992fb6f/jdk/src/share/classes/java/lang/Thread.java#L1003 That's why! Ah what an horror story weekend 🥶 |
Ah, it's going through a java<->native transition. So, a slowdown makes sense. |
normanmaurer
approved these changes
Apr 23, 2023
trustin
approved these changes
Apr 23, 2023
Co-authored-by: Trustin Lee <t@motd.kr>
|
@chrisvest thanks a lot! |
normanmaurer
pushed a commit
that referenced
this pull request
Apr 24, 2023
Motivation: To avoid reference processing overhead while at the same time ensuring that we clear Recycler data structures that were used by now-terminated threads, we need to check if the LocalPool owner thread has terminated. We used to do this by calling `Thread.getState`, but this turns out to be quite expensive on OpenJ9, because they need to halt the target thread before its state can be safely inspected. See the `haltThreadForInspection` call in [`Java_java_lang_Thread_getStateImpl`](https://github.com/eclipse-openj9/openj9/blob/55b9267740559406b8471000cd3a3b6f681dc095/runtime/jcl/common/thread.cpp#L76-L102). This turns out to be a scalability bottleneck in some deployments. Modification: We can get the same semantics as `Thread.getState() == Thread.State.TERMINATED` by checking `!Thread.isAlive()`. The `isAlive` call is cheap on both [OpenJ9](https://github.com/eclipse-openj9/openj9/blob/03967d8cc24dc1f60168110c9d207d2dff57765a/jcl/src/java.base/share/classes/java/lang/Thread.java#L756-L761) and [OpenJDK](https://github.com/openjdk/jdk/blob/master/src/java.base/share/classes/java/lang/Thread.java#L1798-L1815). Result: We no longer see the scalability bottleneck observed in #13347 (comment) Co-authored-by: Trustin Lee <t@motd.kr>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Motivation:
To avoid reference processing overhead while at the same time ensuring that we clear Recycler data structures that were used by now-terminated threads, we need to check if the LocalPool owner thread has terminated.
We used to do this by calling
Thread.getState, but this turns out to be quite expensive on OpenJ9, because they need to halt the target thread before its state can be safely inspected. See thehaltThreadForInspectioncall inJava_java_lang_Thread_getStateImpl.This turns out to be a scalability bottleneck in some deployments.
Modification:
We can get the same semantics as
Thread.getState() == Thread.State.TERMINATEDby checking!Thread.isAlive().The
isAlivecall is cheap on both OpenJ9 and OpenJDK.Result:
We no longer see the scalability bottleneck observed in #13347 (comment)