[
https://issues.apache.org/jira/browse/CASSANDRA-8897?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14560574#comment-14560574
]
Stefania commented on CASSANDRA-8897:
-------------------------------------
bq. That will teach me to try and get all of my outstanding tasks done in a
race before bed. There are still some fairly problematic characteristics.
First: this looks to largely remove the benefit of re-allocating from partially
freed chunks in the same thread, since the thread will immediately free any
chunk that is not its current one. Secondly, it can result in some fairly weird
characteristics in the global pool, since we can return an almost entirely used
Chunk back into the pool for use by another thread. So we could have a large
number of Chunks that are not helpful popping in and out of the global pool,
eliminating the amortization of the overhead costs. This could also lead to the
temporary aggregation of large quantities of mostly-free chunks at the back of
a LocalPool queue, because they were mostly not-free when it fetched and used
them. So it could have only tiny buffers sliced from these chunks, leaving
perhaps 90% of the memory it occupies unused. If it is a long running operation
(say, compaction) this would be fairly undesirable.
The LocalPool now keeps 3 chunks at most. Nonetheless, the things you mention
still apply. So shall we have any thread recycle any chunk as long as it is
free and not used by a local pool (owner null). I guess this is what your
suggestion was aiming at?
> Remove FileCacheService, instead pooling the buffers
> ----------------------------------------------------
>
> Key: CASSANDRA-8897
> URL: https://issues.apache.org/jira/browse/CASSANDRA-8897
> Project: Cassandra
> Issue Type: Improvement
> Components: Core
> Reporter: Benedict
> Assignee: Stefania
> Fix For: 3.x
>
> Attachments: 9240_test_results.txt,
> snapshot-1431582436640-cpu-backtraces.png,
> snapshot-1431582436640-cpu-calltree-compression-8897.nps,
> snapshot-1431582436640-cpu-calltree-compression-trunk.nps
>
>
> After CASSANDRA-8893, a RAR will be a very lightweight object and will not
> need caching, so we can eliminate this cache entirely. Instead we should have
> a pool of buffers that are page-aligned.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
