[
https://issues.apache.org/jira/browse/ZOOKEEPER-3180?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16698826#comment-16698826
]
Rakesh R commented on ZOOKEEPER-3180:
-------------------------------------
[~lvfangmin], Feature looks pretty interesting and thanks for the patch. I will
try to review it when I get a chance.
I'd like to understand scaling of this cache in the production env. Appreciate
if you could help me, adding more details about the memory usage:
# How much is the java heap size of ZooKeeper server and the possible number
of znodes (am assuming 4MB data per znode).
# Based on your experiment, what would be the ideal number of elements to be
kept into the cache for a better read performance gain.
Since you have mentioned that the intention of read cache feature is to reduce
the GC overhead, I would also like to explore the option of enabling the cache
backed by {{off heap}} memory. Again, you have {{4MB}} data in a single zNode
and it would occupy good amount of {{on heap}} memory area if the number of
zNode grows(in future). The off heap implementation uses {{DirectByteByffers}}
to manage cache outside of the JVM heap and provides scalability to a large
memory sizes without GC overhead.
One option could be to provide different cache modes (onheap or offheap) to the
users to efficiently utilize the heap area. Does this make sense to you?
> Add response cache to improve the throughput of read heavy traffic
> -------------------------------------------------------------------
>
> Key: ZOOKEEPER-3180
> URL: https://issues.apache.org/jira/browse/ZOOKEEPER-3180
> Project: ZooKeeper
> Issue Type: Improvement
> Components: server
> Reporter: Fangmin Lv
> Assignee: Brian Nixon
> Priority: Minor
> Labels: pull-request-available
> Fix For: 3.6.0
>
> Time Spent: 3h
> Remaining Estimate: 0h
>
> On read heavy use case with large response data size, the serialization of
> response takes time and added overhead to the GC.
> Add response cache helps improving the throughput we can support, which also
> reduces the latency in general.
> This Jira is going to implement a LRU cache for the response, which shows
> some performance gain on some of our production ensembles.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
