[
https://issues.apache.org/jira/browse/CASSANDRA-15392?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17021455#comment-17021455
]
Blake Eggleston commented on CASSANDRA-15392:
---------------------------------------------
Thanks for the feedback. I just pushed up some commits addressing most points.
Getting rid of linked list queues is a good idea, but don’t think
TinyThreadLocalPool will be sufficient in some cases. If {{MergeIterator#get}}
is called with more than than {{DEFAULT_SIZE}} iterators, we’ll nest pooled
iterators. In cases where there _are_ that many sstables to merge the node is
usually also under significant gc pressure, so it would not be a great time to
abandon this optimization. An intrusive linked stack is the best I could come
up with. It's unbounded, doesn't do any allocations, and has very simple logic,
which was important for preventing oneToOne performance from regressing too
much.
@Jordan, short circuiting the merge iterator in that case is a really good
idea. However, FQL will need a bit of (probably straightforward) rework before
{{MergeIterator#get}} can return a closeable iterator, since it relies on
having separate input and output types. As this relates to SASI, this wouldn’t
be a regression right? I'd guess we're already wasting TrivialOneToOne
iterators?
I’ve also put together a very simple jmh benchmark. After a few iterations of
additional optimization, the manyToOne path saw a 15% increase in throughput in
most cases, and the oneToOne path saw a 2-6% reduction in most cases. Numbers
below:
{code:java}
Baseline:
Benchmark Mode Cnt
Score Error Units
MergeIteratorGetBench.manyToOne sample 59417991
418.721 ± 22.706 ns/op
MergeIteratorGetBench.manyToOne:manyToOne·p0.00 sample
101.000 ns/op
MergeIteratorGetBench.manyToOne:manyToOne·p0.50 sample
148.000 ns/op
MergeIteratorGetBench.manyToOne:manyToOne·p0.90 sample
171.000 ns/op
MergeIteratorGetBench.manyToOne:manyToOne·p0.95 sample
176.000 ns/op
MergeIteratorGetBench.manyToOne:manyToOne·p0.99 sample
188.000 ns/op
MergeIteratorGetBench.manyToOne:manyToOne·p0.999 sample
1456.000 ns/op
MergeIteratorGetBench.manyToOne:manyToOne·p0.9999 sample
9171.213 ns/op
MergeIteratorGetBench.manyToOne:manyToOne·p1.00 sample
38862848.000 ns/op
MergeIteratorGetBench.one sample 53514991
83.452 ± 4.005 ns/op
MergeIteratorGetBench.one:one·p0.00 sample
47.000 ns/op
MergeIteratorGetBench.one:one·p0.50 sample
73.000 ns/op
MergeIteratorGetBench.one:one·p0.90 sample
90.000 ns/op
MergeIteratorGetBench.one:one·p0.95 sample
95.000 ns/op
MergeIteratorGetBench.one:one·p0.99 sample
103.000 ns/op
MergeIteratorGetBench.one:one·p0.999 sample
119.000 ns/op
MergeIteratorGetBench.one:one·p0.9999 sample
3830.003 ns/op
MergeIteratorGetBench.one:one·p1.00 sample
21987328.000 ns/op
Pooled:
Benchmark Mode Cnt
Score Error Units
MergeIteratorGetBench.manyToOne sample 100352304
133.534 ± 1.662 ns/op
MergeIteratorGetBench.manyToOne:manyToOne·p0.00 sample
86.000 ns/op
MergeIteratorGetBench.manyToOne:manyToOne·p0.50 sample
128.000 ns/op
MergeIteratorGetBench.manyToOne:manyToOne·p0.90 sample
143.000 ns/op
MergeIteratorGetBench.manyToOne:manyToOne·p0.95 sample
148.000 ns/op
MergeIteratorGetBench.manyToOne:manyToOne·p0.99 sample
157.000 ns/op
MergeIteratorGetBench.manyToOne:manyToOne·p0.999 sample
172.000 ns/op
MergeIteratorGetBench.manyToOne:manyToOne·p0.9999 sample
3744.000 ns/op
MergeIteratorGetBench.manyToOne:manyToOne·p1.00 sample
13008896.000 ns/op
MergeIteratorGetBench.one sample 97007239
87.340 ± 3.063 ns/op
MergeIteratorGetBench.one:one·p0.00 sample
51.000 ns/op
MergeIteratorGetBench.one:one·p0.50 sample
78.000 ns/op
MergeIteratorGetBench.one:one·p0.90 sample
92.000 ns/op
MergeIteratorGetBench.one:one·p0.95 sample
97.000 ns/op
MergeIteratorGetBench.one:one·p0.99 sample
106.000 ns/op
MergeIteratorGetBench.one:one·p0.999 sample
117.000 ns/op
MergeIteratorGetBench.one:one·p0.9999 sample
3600.000 ns/op
MergeIteratorGetBench.one:one·p1.00 sample
27164672.000 ns/op
{code}
> Pool Merge Iterators
> --------------------
>
> Key: CASSANDRA-15392
> URL: https://issues.apache.org/jira/browse/CASSANDRA-15392
> Project: Cassandra
> Issue Type: Sub-task
> Components: Local/Compaction
> Reporter: Blake Eggleston
> Assignee: Blake Eggleston
> Priority: Normal
> Fix For: 4.0
>
>
> By pooling merge iterators, instead of creating new ones each time we need
> them, we can reduce garbage on the compaction and read paths under relevant
> workloads by ~4% in many cases.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
