We are using CDH 4.3 with Cloudera Manager 4.6.1. We've setup our HBase
cluster giving plenty of RAM to RegionServer, that is setting *Java Heap
Size of HBase RegionServer in Bytes* to 6 GiB. First performance tests
showed really bad results, which we narrowed down to GC problems: the minor
collection was called 10 times a second! Further investigations showed us
that the size of Eden Space was just 6 MiB, which we think can be due to
the reason Tenured Space doesn't grow fast enough from the initial settings.
The good practice for Java long-time running server process is to set *Xms*equal to
*Xmx* to get rid of pauses for resizing, have you considered that?