Anoop Sam John commented on HBASE-16408:

As per the tests so far, even the new off heap BC itself will be faster than L1 
cache specially on high load.   This is mainly due to the fact that we can work 
with very less xmx then and as there are no copies, the GC activity will be 
much lesser.  In case of on heap BC, the issue is our working set will be 
always high. So configuring the IHOP for G1GC will be tricky.   Working set 
being so high will be GC problematic as per the tests.  So what I feel is on 
heap BC should not be used at all.  Specially now we have an off heap version 
which can be almost equal to L1 cache in lower workload and faster when high 
Ya mmaped is considered file mode way only.  We will do a copy on the HFile 
block on hit.  We have to do this also..  This block bytes will be read so many 
time in the read stack (Cell compares etc etc)..  If we get more page faults, 
we will have horrible results.  Thoughts boss?
I plan to open a jira to remove on heap BC support.

> Handle On heap BucketCache size when HeapMemoryManager tunes memory
> -------------------------------------------------------------------
>                 Key: HBASE-16408
>                 URL: https://issues.apache.org/jira/browse/HBASE-16408
>             Project: HBase
>          Issue Type: Sub-task
>            Reporter: Anoop Sam John
>            Assignee: Anoop Sam John
>             Fix For: 2.0.0

This message was sent by Atlassian JIRA

Reply via email to