In my cluster, I changed keepAliveTime from 60 s to 3600 s. Increasing RES is slowed down.
Why increasing keepAliveTime of HBase thread pool is slowing down our problem occurance [RES value increase]?
You can go through the source of sun.nio.ch.Util. Every thread hold 3 softreference of direct buffer(mustangsrc) for reusage. The code names the 3 softreferences buffercache. If the buffer was all occupied or none was suitable in size, and new request comes, new direct buffer is allocated. After the service, the bigger one replaces the smaller one in buffercache. The replaced buffer is released.
So I think we can add a parameter to change keepAliveTime of Htable thread pool.