Implement BlockCache using custom LRU/ARC mechanism rather than SoftReferences
------------------------------------------------------------------------------
Key: HBASE-965
URL: https://issues.apache.org/jira/browse/HBASE-965
Project: Hadoop HBase
Issue Type: New Feature
Affects Versions: 0.19.0
Reporter: Jonathan Gray
Priority: Minor
Fix For: 0.19.0
BlockCaching is now turned on by default as of HBASE-953. This improves
all-around performance but is very taxing on the GC. Random reading tends to
thrash the cache.
As discussed, rather than rely on JVM LRU GC, we should implement our own
tracking of cached blocks using an ARC-like algorithm. This will also allow
users to set the cache size per-table or per-family rather than just taking
from the allotted Heap pool.
Paper on ARC algorithm (a scan-resistant LRU):
http://www.almaden.ibm.com/cs/people/dmodha/ARC.pdf
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
