You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@ignite.apache.org by "Kirill Gusakov (Jira)" <ji...@apache.org> on 2022/08/08 00:15:00 UTC
[jira] [Commented] (IGNITE-17423) Investigate performance of insert only YCSB benchmark
[ https://issues.apache.org/jira/browse/IGNITE-17423?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17576463#comment-17576463 ]
Kirill Gusakov commented on IGNITE-17423:
-----------------------------------------
I pushed the localized benchmark to the [https://github.com/gridgain/apache-ignite-3/tree/ignite-17423] and create some custom JFR events for hot methods, which I found:
* we have significant differences in the performance of key-value and JDBC API again, created the separate issue with details https://issues.apache.org/jira/browse/IGNITE-17482
* in the key-value benchmark itself: we spend half of the time (280us of 528us) for appending the entries to the RAFT log storage (RocksDB at the current time) with ENABLED FSYNC. We need fsync per record at the time, and can't disable it (look at the charts, considering that we have 2 append entries per insert query)
* in key-value benchmark: we spent near the 1/5 (120us of 528us) part of the time to marshall/unmarshall (look at the charts, considering that we have 2 marshall and 2 unmarshalls per insert query)
> Investigate performance of insert only YCSB benchmark
> -----------------------------------------------------
>
> Key: IGNITE-17423
> URL: https://issues.apache.org/jira/browse/IGNITE-17423
> Project: Ignite
> Issue Type: Task
> Reporter: Kirill Gusakov
> Assignee: Kirill Gusakov
> Priority: Major
> Labels: ignite-3
> Attachments: Screenshot 2022-08-08 at 03.12.31.png, Screenshot 2022-08-08 at 03.12.46.png, Screenshot 2022-08-08 at 03.13.10.png
>
>
> At the moment performance results of insert-only YCSB bencmark from https://issues.apache.org/jira/browse/IGNITE-17220 and from my local run of synthetic put/insert handmade benchmarks looks so different - I have near the 2000k inserts per second vs the 200 in the YCSB benchmark (on the same my machine).
>
> Needed further investigations if it can be described by:
> * jdbc client penalties (not covered by synthetic put only benchmarks)
> * remote machine latency
> * <any other reasons>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)