[ 
https://issues.apache.org/jira/browse/IGNITE-24118?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17921512#comment-17921512
 ] 

Vladislav Pyatkov commented on IGNITE-24118:
--------------------------------------------

I run the test many times and analyzed every commit between these revisions:
* Empty space of the hard disk influences to the test result (the best result 
when the disk is used less than half).
* Size of lock table also have a matter of the test performance (after this 
thicket IGNITE-24267 we do not pre-allocate lock table).
* Using Objects#hash is time consuming. In particular when all parameters are 
primitives (like for TablePartitionId#hashCode) better to implement internal 
hash calculation.
!screenshot-1.png! 

> Investigate performance drop on master.
> ---------------------------------------
>
>                 Key: IGNITE-24118
>                 URL: https://issues.apache.org/jira/browse/IGNITE-24118
>             Project: Ignite
>          Issue Type: Improvement
>            Reporter: Alexey Scherbakov
>            Assignee: Vladislav Pyatkov
>            Priority: Major
>              Labels: ignite-3
>         Attachments: screenshot-1.png
>
>
> I observe stable performance drop on recent main on UpsertKvBenchmark.
> Benchmark settings:
> 32 threads, 32 partitions, AI persist engine, 20G pagecache size
> System.setProperty("LOGIT_STORAGE_ENABLED", "true");
> System.setProperty(IgniteSystemProperties.IGNITE_SKIP_REPLICATION_IN_BENCHMARK,
>  "false");
> System.setProperty(IgniteSystemProperties.IGNITE_SKIP_STORAGE_UPDATE_IN_BENCHMARK,
>  "false");
> Results:
> rev 32737c0dc9fcd0632ba37e2949a40b199429fddb
> Benchmark (batch) (fsync) (partitionCount) Mode Cnt Score Error Units
> UpsertKvBenchmark.upsert 1 false 32 thrpt 20 181908.090 ± 26821.026 ops/s
> rev e41ee1dd796c209115d818a26aded1dbf27bd9a1 (recent)
> Benchmark                 (batch)  (fsync)  (partitionCount)   Mode  Cnt      
>  Score       Error  Units
> UpsertKvBenchmark.upsert        1    false                32  thrpt   20  
> 173093.148 ± 29217.456  ops/s
> The drop is even bigger on a branch ignite-23304 after merging with recent 
> main.
> This requires investigation.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

Reply via email to