[
https://issues.apache.org/jira/browse/HBASE-27683?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Xiaolin Ha updated HBASE-27683:
-------------------------------
Attachment: image-2023-05-29-17-16-55-133.png
> Should support single call queue mode for RPC handlers while separating by
> request type
> ---------------------------------------------------------------------------------------
>
> Key: HBASE-27683
> URL: https://issues.apache.org/jira/browse/HBASE-27683
> Project: HBase
> Issue Type: Improvement
> Components: Performance, rpc
> Affects Versions: 2.5.3
> Reporter: Xiaolin Ha
> Assignee: Xiaolin Ha
> Priority: Major
> Attachments: image-2023-05-29-17-16-55-133.png
>
>
> Currently we not only seperate call queues by request type, e.g. read, write,
> scan, but also distinguish queues for handlers by the config
> `hbase.ipc.server.callqueue.handler.factor`, whose description is as follows,
> {code:java}
> Factor to determine the number of call queues.
> A value of 0 means a single queue shared between all the handlers.
> A value of 1 means that each handler has its own queue. {code}
> But I think what we want is not only one queue for all the requests, or each
> handler has its own queue. We also want each request type has its own queue.
> Distinguishing queues in the same type of requests will make some handlers
> too iddle but some handlers too busy under current balanced/random RPC
> executor framework. For the extrem case, each handler has its own queue, then
> if a large request comes for a handler, duing to he executor dispath calls
> without considering the queue size or the state of the handler, the
> afterwards coming requests will be queued until the handler complete the
> large slow request. While other handlers may process small requests quickly,
> but they can not help or grab calls from the busy queue, they must stay and
> wait it own queue jobs coming. Then we can see the queue time of some
> requests are long but there are iddle handlers.
> We can also see these circumstances, that the queue time of calls is too
> larger than the process time, sometimes twice or more. Restarting the slow RS
> will make these problems disappear.
> By using single call queue for each request type, we can fully use the
> handler resources.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
