[
https://issues.apache.org/jira/browse/HDFS-6599?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14042481#comment-14042481
]
Kihwal Lee commented on HDFS-6599:
----------------------------------
This is caused by a change in {{BlockPlacementPolicyDefault#isGoodTarget()}}.
The synchronization overhead in {{HeartBeatManager}} or {{DatanodeManager}}
might also be worse.
In 0.23 it didn't go there at all. It just looked at the topology.
{{int size = clusterMap.getNumOfLeaves()}}
In 2.4.x, it is calling this:
{{int size = stats.getNumDatanodesInService()}}
Since the network topology doesn't tell you the accurate state of the cluster,
it is a right thing to do semantically. But its performance impact should have
been considered. This change was made in HDFS-5837.
> 2.4 addBlock is 10 to 20 times slower compared to 0.23
> ------------------------------------------------------
>
> Key: HDFS-6599
> URL: https://issues.apache.org/jira/browse/HDFS-6599
> Project: Hadoop HDFS
> Issue Type: Bug
> Affects Versions: 2.3.0, 2.4.0
> Reporter: Kihwal Lee
> Priority: Critical
>
> From one of our busiest 0.23 clusters:
> {panel}
> "AddBlockAvgTime" : 0.9514711501719515
> "CreateAvgTime" : 1.7564162389174
> "CompleteAvgTime" : 1.3310406035056548
> "BlockReceivedAndDeletedAvgTime" : 0.661210005151392
> {panel}
> From a not-so-busy 2.4 cluster:
> {panel}
> "AddBlockAvgTime" : 10.083333333333334
> "CreateAvgTime" : 1.0
> "CompleteAvgTime" : 1.1111111111111112
> "BlockReceivedAndDeletedAvgTime" : 0.07692307692307694
> {panel}
> When the 2.4 cluster gets a moderate amount of write requests, the latency is
> terrible. E.g. addBlock goes upward of 60ms.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
