[
https://issues.apache.org/jira/browse/HIVE-16992?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16393968#comment-16393968
]
Sergey Shelukhin commented on HIVE-16992:
-----------------------------------------
Looks like lambda needs to be reduced massively, right now it is indeed a
glorified LRU.
Overall it seems like due to the clock being the counter of cache ops, the
LRU/LFU balance is also dependent to some extent on cache size.
I will post a patch next week after clean up and looking at some query runs.
> LLAP: better default lambda for LRFU policy
> -------------------------------------------
>
> Key: HIVE-16992
> URL: https://issues.apache.org/jira/browse/HIVE-16992
> Project: Hive
> Issue Type: Bug
> Reporter: Sergey Shelukhin
> Assignee: Sergey Shelukhin
> Priority: Major
>
> LRFU is currently skewed heavily towards LRU; there are 10k-s or 100k-s of
> buffers tracked during a typical workload, but the heap size is around 700.
> We should see if making it closer to LFU (by tweaking the lambda) will
> improve hit rate with small queries infrequently interleaved with large
> scans; and whether it will have negative effects due to perf overhead.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
