[
https://issues.apache.org/jira/browse/HBASE-17204?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16258420#comment-16258420
]
Anoop Sam John commented on HBASE-17204:
----------------------------------------
We fixed critical issues around BucketCache including the
fragmentation one. Still the open one is we maintain at least one free
bucket for every size. This gives some wastage of space when the
bucket sizes are not properly tuned. That sub jira is still open !
May be we should just allow all the buckets of a size to be converted
to other size as needed? No reserve of one block at least. U asked
this Q some where.
On the route to make off heap BC def ON , we should
1. Remove the L1, L2 concept fully. By default yes this is the way it
works now also. When BC is ON, by default we will keep data blocks in
BC only and index blocks in LRU cache. We have a config to make the BC
a strict L2 cache (By L2 as a victim handler for L1)
2. Dont keep BC as a victim handler for LRU cache at all. Right now
this is done when combined mode is true/false. Lets keep it simple.
Index blocks will go to LRU and data blocks to BC. When LRU is not
enough, we will see evictions there.
3. Reduce the heap % for LRU. Right now it degfaults to 40%. But when
it is going to keep index blocks alone, lets make the default to be
10%?
4. Off heap mode BC is ON by def. We need a def off heap size.
Remember the off heap size is not given as a % of some thing. It is
exact value. May be we should give a def value of 8 GB? (or much lesser or
greater?)
> Make Bucket Cache off heap cache default ON
> -------------------------------------------
>
> Key: HBASE-17204
> URL: https://issues.apache.org/jira/browse/HBASE-17204
> Project: HBase
> Issue Type: Improvement
> Affects Versions: 2.0.0
> Reporter: Anoop Sam John
> Assignee: Anoop Sam John
> Priority: Blocker
> Fix For: 2.0.0-beta-1
>
>
> L2 cache can be used for data blocks. By default it is off now. After
> HBASE-11425 work, L2 off heap cache can equally perform with L1 on heap
> cache. On heavy loaded workload, this can even out perform L1 cache. Pls see
> recently published report by Alibaba. Also this work was backported by
> Rocketfuel and similar perf improvement report from them too.
> Let us turn L2 off heap cache ON. As it is off heap, we can have much larger
> sized L2 BC. What should be the default size?
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
