[
https://issues.apache.org/jira/browse/IGNITE-6699?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16358296#comment-16358296
]
Ivan Fedotov commented on IGNITE-6699:
--------------------------------------
[~vozerov],
I inserted a collection and now data from addData(key,value) method are being
cached in list [1], results of benchmark on adding 1000 integers seem good:
Benchmark Mode Cnt
Score Error Units
JmhStreamerCollectionBenchmark.addData avgt 21 *332,170* ± 34,122
us/op
JmhStreamerKeyValueBenchmark.addData avgt 21 *341,228* ± 23,528 us/op
But some details of implementation are still unclear to me:
1. Is default collection size of 500 will be enough? I took this number
slightly less than default buffer size for node
(IgniteDataStreamer.DFLT_PER_NODE_BUFFER_SIZE).
2. If I add some data per key value, for example 100 values (< default buffer
size) and after that add collection also with 100 values (< default buffer size
per node), should I catch data from both operations in collection or only in
case of streaming per key/value? This question relates to other combinations of
these methods.
3. What did you mean by saying "from one thread"? Is placing collection to
synchronize block will be enough?
[1]https://github.com/apache/ignite/pull/3442
> Optimize client-side data streamer performance
> ----------------------------------------------
>
> Key: IGNITE-6699
> URL: https://issues.apache.org/jira/browse/IGNITE-6699
> Project: Ignite
> Issue Type: Task
> Components: streaming
> Affects Versions: 2.3
> Reporter: Vladimir Ozerov
> Assignee: Ivan Fedotov
> Priority: Major
> Labels: iep-1, performance
> Fix For: 2.5
>
>
> Currently if a user has several server nodes and a single client node with
> single thread pushing data to streamer, he will not be able to load data at
> maximum speed. On the other hand, if he start several data loading threads,
> throughput will increase.
> One of root causes of this is bad data streamer design. Method
> {{IgniteDataStreamer.addData(K, V)}} returns new feature for every operation,
> this is too fine grained approach. Also it generates a lot of garbage and
> causes contention on streamer internals.
> Proposed implementation flow:
> 1) Compare performance of {{addData(K, V)}} vs {{addData(Collection)}}
> methods from one thread in distributed environment. The latter should show
> considerably higher throughput.
> 2) Users should receive per-batch features, rather than per-key.
> 3) Try caching thread data in some collection until it is large enough to
> avoid contention and unnecessary allocations.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
