[ https://issues.jboss.org/browse/ISPN-6494?page=com.atlassian.jira.plugin.... ] Bela Ban edited comment on ISPN-6494 at 8/12/19 6:33 AM: --------------------------------------------------------- [~johnou] No updates. I want to complete the last JIRA(s) in 4.1.3, then release it and get back to 5.0 work. After that release, I'm willing to tackle this. {{no-bundler}} has _never_ shown up as top performer in _any_ of my perftests! {{transfer-queue}} has always been near the top, and the lockless- and ringbuffer-based bundlers have been the fastest bundlers around. If you send many messages (ie. in seperate threads) to the same destinations (including {{null}}), then bundling/batching will get you much better performance, as fewer resources such as threads and locks will be used. Even for OOB messages, which have a different thread dispatching policy, handling a batch of 10 is much more efficient than handling 10 separate messages. Without knowing your specific access patterns, I venture to say that batching will greatly increase your perf. Of course, this depends on * How many threads are sending messages * To which destinations * OOB or regular messages * Configured message processing policy * Delivery at the application (blocking versus non-blocking delivery) and processing time Do you have a specific perftest mimicking your application? That would be best to tune perf. On the JGroups side, we have UPerf and MPerf as part of JGroups (plus IspnPerfTest as separate project on GH). was (Author: belaban): [~johnou] No updates. I want to complete the last JIRA(s) in 4.1.3, then release it and get back to 5.0 work. After that release, I'm willing to tackle this. {{no-bundler}} has _never_ shown up as top performer in _any_ of my perftests! {{transfer-queue}} has always been near the top, and the lockless- and ringbuffer-based bundler have been the fastest bundlers around. If you send many messages (ie. in seperate threads) to the same destinations (including {{null}}), then bundling/batching will get you much better performance, as fewer resources such as threads and locks will be used. Even for OOB messages, which have a different thread dispatching policy, handling a batch of 10 is much more efficient than handling 10 separate messages. Without knowing your specific access patterns, I venture to say that batching will greatly increase your perf. Of course, this depends on * How many threads are sending messages * To which destinations * OOB or regular messages * Configured message processing policy * Delivery at the application (blocking versus non-blocking delivery) and processing time Do you have a specific perftest mimicking your application? That would be best to tune perf. On the JGroups side, we have UPerf and MPerf as part of JGroups (plus IspnPerfTest as separate project on GH). -- This message was sent by Atlassian Jira (v7.12.1#712002)