[
https://issues.apache.org/jira/browse/CASSANDRA-5661?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13704045#comment-13704045
]
Pavel Yaskevich edited comment on CASSANDRA-5661 at 7/10/13 12:54 AM:
----------------------------------------------------------------------
I have started working on integrating multiway pool and it looks like we have
two problems:
#1. As each SegmentedFile has to return unique instance using
"createReader(String)" LoadingCache won't do for us, as we need get(K,
Callable) per SSTableReader.
#2. as MultiWay returns a handle I changed RAR to have a setHandle method
instead of passing SegmentedFile into constructor, which seems a bit hacky to
me as we need to be careful in maintaining that relationship...
I did some performance testing (with attached patch) where MultiwayPool
allocated per instance because we can't specify loader in borrow(...) yet,
which should be a best case for it, not in terms of memory usage but
contention. I loaded 5,000,000 keys with following stress command
(./tools/bin/cassandra-stress -n 5000000 -S 512 -C 20 -Z
LeveledCompactionStrategy) for initial data and then I made it run in a loop
and was doing reads in parallel.
With writes:
Average read performance for MultiwayPool: median 6.2, 95th 11.4, 99.9th 78.8
Average read performance for FileCacheService: median: 5.3, 95th 9.6, 99.9th
73.1
No writes, no compaction:
Average read performance for MultiwayPool: median 2.3, 95th 3.2, 99.9th 21.3
Average read performance for FileCacheService: median: 1.7, 95th 2.9, 99.9th
19.2
I tried doing range_slice but due to timeouts I couldn't really complete test
on any of the implementations, median latenties on average different by 3-4 ms.
Edit: I forgot to mention that I hardcoded maxSize per MultiwayPool instance
which was fine for that test, but we really need a way to weight items if we
are going to use it globally.
was (Author: xedin):
I have started working on integrating multiway pool and it looks like we
have two problems:
#1. As each SegmentedFile has to return unique instance using
"createReader(String)" LoadingCache won't do for us, as we need get(K,
Callable) per SSTableReader.
#2. as MultiWay returns a handle I changed RAR to have a setHandle method
instead of passing SegmentedFile into constructor, which seems a bit hacky to
me as we need to be careful in maintaining that relationship...
I did some performance testing (with attached patch) where MultiwayPool
allocated per instance because we can't specify loader in borrow(...) yet,
which should be a best case for it, not in terms of memory usage but
contention. I loaded 5,000,000 keys with following stress command
(./tools/bin/cassandra-stress -n 5000000 -S 512 -C 20 -Z
LeveledCompactionStrategy) for initial data and then I made it run in a loop
and was doing reads in parallel.
With writes:
Average read performance for MultiwayPool: median 6.2, 95th 11.4, 99.9th 78.8
Average read performance for FileCacheService: median: 5.3, 95th 9.6, 99.9th
73.1
No writes, no compaction:
Average read performance for MultiwayPool: median 2.3, 95th 3.2, 99.9th 21.3
Average read performance for FileCacheService: median: 1.7, 95th 2.9, 99.9th
19.2
I tried doing range_slice but due to timeouts I couldn't really complete test
on any of the implementations, median latenties on average different by 3-4 ms.
> Discard pooled readers for cold data
> ------------------------------------
>
> Key: CASSANDRA-5661
> URL: https://issues.apache.org/jira/browse/CASSANDRA-5661
> Project: Cassandra
> Issue Type: Bug
> Components: Core
> Affects Versions: 1.2.1
> Reporter: Jonathan Ellis
> Assignee: Pavel Yaskevich
> Fix For: 2.0
>
> Attachments: CASSANDRA-5661-multiway-per-sstable.patch,
> CASSANDRA-5661.patch, DominatorTree.png, Histogram.png
>
>
> Reader pooling was introduced in CASSANDRA-4942 but pooled
> RandomAccessReaders are never cleaned up until the SSTableReader is closed.
> So memory use is "the worst case simultaneous RAR we had open for this file,
> forever."
> We should introduce a global limit on how much memory to use for RAR, and
> evict old ones.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
