[
https://issues.apache.org/jira/browse/GEODE-7245?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17290531#comment-17290531
]
ASF GitHub Bot commented on GEODE-7245:
---------------------------------------
pivotal-jbarrett commented on a change in pull request #6037:
URL: https://github.com/apache/geode/pull/6037#discussion_r582368198
##########
File path:
geode-membership/src/main/java/org/apache/geode/distributed/internal/membership/gms/GMSMembership.java
##########
@@ -1231,23 +1197,16 @@ public void shutdownMessageReceived(ID id, String
reason) {
if (logger.isDebugEnabled()) {
logger.debug("Membership: recording shutdown status of {}", id);
}
- synchronized (this.shutdownMembers) {
- this.shutdownMembers.put(id, id);
- services.getHealthMonitor()
- .memberShutdown(id, reason);
- services.getJoinLeave().memberShutdown(id, reason);
- }
+ this.shutdownMembers.add(id);
+ services.getJoinLeave().memberShutdown(id, reason);
}
@Override
public Set<ID> getMembersNotShuttingDown() {
- latestViewReadLock.lock();
- try {
- return latestView.getMembers().stream().filter(id ->
!shutdownMembers.containsKey(id))
+ synchronized (shutdownMembers) {
Review comment:
If the `shutdownMembers` list isn't changing frequently, like every ms,
then repeatedly recreating the same set is just adding pressure to the GC and
CPU time all while under a synchronized lock for no real gain. I assume this
`shutdownMembers` list is updated with the member id of the member that has
said it is or has gone down, which isn't a frequent event in production. Seems
like we should be optimizing this method to be returning a rather static list
of members known not to be shutting down even if that means the infrequent
setting of a shutting down member takes a bit more time to handle, by copying
lists and invalidating or updating the not shutdown servers list.
At the very least the addition of this synchronized block is going to have a
noticeable impact on the p2p messaging at scale since all messages will be
serialized through this lock. If we deem refactoring out the list creation is
out of scope for this PR then using a read/write lock here should be in scope
as a first step.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
> Remove most uses of latestViewReadLock in GMSMembershipManager
> --------------------------------------------------------------
>
> Key: GEODE-7245
> URL: https://issues.apache.org/jira/browse/GEODE-7245
> Project: Geode
> Issue Type: Improvement
> Components: membership
> Reporter: Bruce J Schuchardt
> Priority: Major
> Labels: pull-request-available
>
> Changes for GEODE-7196 made the latestView immutable so there is no reason,
> in most cases, to use the latestViewReadlock when accessing the view. A
> simple volatile read of the field & storing the value in a method variable
> can replace most uses of the latestViewReadLock, especially those uses that
> are in the path of sending or receiving messages.
> Performance could be measured with our benchmarks to see if this change helps.
>
> latestViewReadLock is currently used in getCoordinator(), memberExists(),
> getMembersNotShuttingDown(), getAllMembers(), isShunnedOrNew(),
> isSurpriseMember(), doWithViewLocked().
> doWithViewLocked() is used to add membership listeners, and we don't want the
> view to change until this operation completes. Therefore, we should keep the
> lock in this method.
> getCoordinator(), memberExists(), getMembersNotShuttingDown(),
> getAllMembers() use latestViewReadLock to access latestView. In these methods
> we can replace the lock with a volatile read of the latestView and storing
> the value in a method variable.
> isShunnedOrNew() accesses latestView, and 2 ConcurrentHashMaps -
> shunnedMembers and surpriseMembers. Since the hashmaps are concurrent and
> latestView is unmodifiable, a volatile read could replace the lock. Same for
> isSurpriseMember(), which uses surpriseMembers hashmap.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
