I've been running OF 3.7.1 with 1.7u5 for a long time in a two node configuration with Coherence for clustering. I recently upgraded our DR environment to 3.8.1 from SVN and replaced Coherence with Hazelcast. With 1.7u5 it works great - It was running for 36 hours with 2 users connected and CPU was basically flat.
With either of the current JREs 1.6u39 or 1.7u5 I see periodic high CPU on both nodes. Typically the CPU utilization lasts for a minute or so, however it corresponds with an exception in org.jivesoftware.openfire.nio.ConnectionHandler on the other node. Both nodes seem to periodically experience this exception due to SSL padding issues from health checks from our load balancers (which is really just XMPPHP connecting and authenticating using TLS). Even just running for a couple of hours I see spikes up to 10% CPU.
I've seen lots of people complaining about SSL padding errors on Java 1.7, but never any real solutions. Has anyone experienced similar performance anolomies with Openfire and Java 1.7?