[
https://issues.apache.org/jira/browse/HBASE-12609?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14233263#comment-14233263
]
Lars Hofhansl commented on HBASE-12609:
---------------------------------------
bq. the replication will wait the current ship returned to do next ship
Yep. RPC in this way is notoriously bad in keeping the latency X bandwidth of
the network utilized.
> Consider bandwidth throttling in
> ReplicationSource.readAllEntriesToReplicateOrNextFile
> --------------------------------------------------------------------------------------
>
> Key: HBASE-12609
> URL: https://issues.apache.org/jira/browse/HBASE-12609
> Project: HBase
> Issue Type: Improvement
> Components: Replication
> Affects Versions: 0.98.8
> Reporter: cuijianwei
> Priority: Minor
>
> The push size in ReplicationSource.shipEdits is controlled by
> 'replicationQueueSizeCapacity' and 'replicationQueueNbCapacity'. When the
> push size is much bigger than throttler's bandwidth(replication throttler is
> enabled), ReplicationSource will have a outgoing bandwidth peak followed by a
> period of sleep controlled by the throttler. How about considering bandwidth
> throttling in ReplicationSource.readAllEntriesToReplicateOrNextFile? such as
> the following code:
> {code}
> // Stop if too many entries or too big
> long bandwidth =
> this.conf.getLong("replication.source.per.peer.node.bandwidth", 0);
> if ((bandwidth > 0 && currentSize > bandwidth) || // ==> also consider
> bandwidth throttling
> currentSize >= this.replicationQueueSizeCapacity ||
> entries.size() >= this.replicationQueueNbCapacity) {
> break;
> }
> {code}
> Then, ReplicationSource will have more steady outgoing bandwidth.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
