Thread BLOCKED issues of DestinationCache in DefaultSubscriptionRegistry
See original GitHub issueWe are using the spring-messaging to implement a STOMP server (using SimpleBrokerMessageHandler). The Client will subscribe on 5 channel and everything is ok when there are only a few users. However, when the online user is above ~ 700, the websocket channel is “out of response”.
After analysis, I found many other thread has been “BLOCKED” by DestinationCache, as follows:
"pk-ws-worker-100-thread-78" #560 prio=5 os_prio=0 tid=0x00007ff19c182000 nid=0x1252 waiting for monitor entry [0x00007ff0a8d9f000]
java.lang.Thread.State: BLOCKED (on object monitor)
at org.springframework.messaging.simp.broker.DefaultSubscriptionRegistry$DestinationCache.getSubscriptions(DefaultSubscriptionRegistry.java:269)
- waiting to lock <0x00000004c007ec20> (a org.springframework.messaging.simp.broker.DefaultSubscriptionRegistry$DestinationCache$1)
And part of the code are as follows:
public LinkedMultiValueMap<String, String> getSubscriptions(String destination, Message<?> message) {
LinkedMultiValueMap<String, String> result = this.accessCache.get(destination);
if (result == null) {
synchronized (this.updateCache) {
result = new LinkedMultiValueMap<>();
for (SessionSubscriptionInfo info : subscriptionRegistry.getAllSubscriptions()) {
for (String destinationPattern : info.getDestinations()) {
if (getPathMatcher().match(destinationPattern, destination)) {
for (Subscription sub : info.getSubscriptions(destinationPattern)) {
result.add(info.sessionId, sub.getId());
}
}
}
}
if (!result.isEmpty()) {
this.updateCache.put(destination, result.deepCopy());
this.accessCache.put(destination, result);
}
}
}
return result;
}
As you can see, the code inside synchronized will traverse all subscription, which will cost too much time and block other Thread.
Also, the accessCache / updateCache is not works if the client has not success make the subscription, which will make the situation worse.
We try to increase the cache limit and it does’t work for our case.
To solve the problem, we remove the DestinationCache and reimplement an Map -> <sessionId -> subsId> inside SessionSubscriptionRegistry. (in our own codebase of course)
After theses change, the server can handle more than 5K online users with no problem.
Meanwhile, I noticed that DefaultSubscriptionRegistry and DestinationCache has been there for many years.
So, I just wonder is it ok to make a pr? Or the existing DestinationCache is good for some other reason?
Issue Analytics
- State:
- Created 4 years ago
- Comments:6 (2 by maintainers)
We observed the same. The CPU was BLOCKED on synchronized blocks in DefaultSubscriptionRegistry causing a bottleneck. We were lucky, that we had no subscription pattern matching so we was able to solve it by reimplemeting DefaultSubscriptionRegistry. We used two concurrentMaps and it’s.computeX() methods:
And registered it like this:
It more than doubled the performance of a simple broker.
Apparently you can have mutliple subscriptions with the same subscription-id and this change breaks that requirement. Specifically UserDestinationMessageHandler will reuse the same topic subscription message to create the session specific subscription to that topic, which means that according to which thread gets to insert the subscription first wins.