[
https://issues.apache.org/jira/browse/HBASE-28556?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17843699#comment-17843699
]
Istvan Toth commented on HBASE-28556:
-------------------------------------
I have run somewhat better tests with my fixed PerformanceEvaluationTool
(patches not yet published):
{noformat}
hbase -Xmx2g org.apache.hadoop.hbase.rest.PerformanceEvaluation
--host=ccycloud-1.stoty.root.comops.site:20550 ---enableSsl=false --rows=250
--api=rest_remote --nomapred=true scanRange10000 20
- JDK17, 4GB heap , patch: ~120s wall clock, 303s REST CPU usage
- JDK17, 1GB heap, patch: ~130s wall clock, 340s REST CPU usage
- JDK17, 4GB heap, NOpatch: ~120 wall clock, 360 REST CPU usage
- JDK17, 1GB heap, NOpatch: ~150 wall clock, 405s REST CPU usage
{noformat}
> Reduce memory copying in Rest server when serializing CellModel to Protobuf
> ---------------------------------------------------------------------------
>
> Key: HBASE-28556
> URL: https://issues.apache.org/jira/browse/HBASE-28556
> Project: HBase
> Issue Type: Improvement
> Components: REST
> Reporter: Istvan Toth
> Assignee: Istvan Toth
> Priority: Minor
> Labels: pull-request-available
>
> The REST server does a lot of unneccessary coping, which could be avoided at
> least for protobuf encoding.
> - -It uses ByteStringer to handle ByteBuffer backed Cells. However, it uses
> the client API, so it should never encounter ByteBuffer backed cells.-
> - It clones everything from the cells (sometimes multiple times) before
> serializing to protbuf.
> We could mimic the structure in Cell, with array, offset and length for each
> field, in CellModel and use the appropriate protobuf setters to avoid the
> extra copies.
> There may or may not be a way to do the same for JSON and XML via jax-rs, I
> don't know the frameworks well enough to tell, but if not, we could just do
> the copying in the getters for them, which would not make things worse.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
