[
https://issues.apache.org/jira/browse/IGNITE-22842?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17886340#comment-17886340
]
Vladislav Pyatkov commented on IGNITE-22842:
--------------------------------------------
Local benchmark result:
Benchmark (clusterSize) (fsync) (useSyncOptimization) Mode
Cnt Score Error Units
SelectBenchmark.kvThinGet 1 false true avgt
20 86,590 ± 2,708 us/op
SelectBenchmark.kvThinGet 1 false false avgt
20 94,047 ± 0,979 us/op
The useSyncOptimization is a flag to apply this optimization or not. Based on
this measurement, the optimization gives about an 8% reduction in latency.
> The client returns to the fork-join pool after handling operations on the
> server side
> -------------------------------------------------------------------------------------
>
> Key: IGNITE-22842
> URL: https://issues.apache.org/jira/browse/IGNITE-22842
> Project: Ignite
> Issue Type: Improvement
> Reporter: Vladislav Pyatkov
> Assignee: Vladislav Pyatkov
> Priority: Major
> Labels: ignite-3
>
> h3. Motivation
> In synchronous operation, we return the result to the fork join pool before
> providing the result to the client in their pool. Of course, the tric costs
> several microseconds.
> {noformat}
> loadSchemaAndReadData:ForkJoinPool.commonPool-worker-9 0.2 9950400 9950600
> Here is hidden 5.2 us
> kvClientPutEndMark:org.apache.ignite.internal.benchmark.InsertBenchmark.kvThinInsert-jmh-worker-1
> 0.0 9955800 9955800
> {noformat}
> The issue is simullar IGNITE-22838 but about that type of operation that
> started on the client side.
> h3. Definition of done
> We need to develop a strategy that does not involve the extra thread in
> synchronous operation.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
