You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@hbase.apache.org by "Bryan Beaudreault (Jira)" <ji...@apache.org> on 2023/03/13 15:43:00 UTC

[jira] [Created] (HBASE-27710) ByteBuff ref counting is too expensive for on-heap buffers

Bryan Beaudreault created HBASE-27710:
-----------------------------------------

             Summary: ByteBuff ref counting is too expensive for on-heap buffers
                 Key: HBASE-27710
                 URL: https://issues.apache.org/jira/browse/HBASE-27710
             Project: HBase
          Issue Type: Improvement
            Reporter: Bryan Beaudreault


We recently were upgrading an internal project's hbase client from 1.2 to 2.5. The project does raw HFile.Reader reading of some hfiles. We noticed an over 4x performance regression with the new client. Profiling showed that the majority of the time was spent in SingleByteBuff.checkRefCount. That ref counting is necessary for off-heap reads, but this case was using on-heap. Removing checkRefCount eliminated the regression.

I asked on the dev list and we agreed to remove checkRefCount calls for on-heap buffers: https://lists.apache.org/thread/2qm0g8fbxb0no4ozfx7bw8xmkryt4zz0



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