[jira] [Comment Edited] (HBASE-28485) Re-use ZstdDecompressCtx/ZstdCompressCtx for performance

["Charles Connell (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HBASE-28485?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17834068#comment-17834068 ] 

Charles Connell edited comment on HBASE-28485 at 4/4/24 8:03 PM:
-----------------------------------------------------------------

I've attached a flamegraph showing ZstdDecompressCtx constructor taking 0.81% of RegionServer CPU time, and another case of pathological RegionServer performance where it takes 1.16% of CPU time.


was (Author: charlesconnell):
I've attached a flamegraph showing ZstdDecompressCtx constructor taking 0.81% of RegionServer CPU time

> Re-use ZstdDecompressCtx/ZstdCompressCtx for performance
> --------------------------------------------------------
>
>                 Key: HBASE-28485
>                 URL: https://issues.apache.org/jira/browse/HBASE-28485
>             Project: HBase
>          Issue Type: Improvement
>            Reporter: Charles Connell
>            Assignee: Charles Connell
>            Priority: Major
>         Attachments: async-prof-flamegraph-cpu_event-1712150670836-cpu.html, async-prof-pid-1324144-cpu-1.html
>
>
> The zstd documentation recommends re-using context objects when possible, because their creation has some expense. They can be more cheaply reset than re-created. In ZstdDecompressor and ZstdCompressor, we create a new context object for every call to decompress() and compress(). In CPU profiles I've taken at my company, the constructor of ZstdDecompressCtx can sometimes represent 10-25% of the time spent in zstd decompression, which itself is 5-10% of a RegionServer's total CPU time. Avoiding this performance penalty won't lead to any massive performance boost, but is a nice little win.



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