[ https://issues.apache.org/jira/browse/HDFS-11461?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15886922#comment-15886922 ]
Arpit Agarwal edited comment on HDFS-11461 at 2/28/17 12:40 AM: ---------------------------------------------------------------- [~hanishakoneru], a few comments: # {{DataNode#shutdown}} should use slowDiskDetectionThread.join() instead of sleep. # {{diskOutliers}} should maintain the mean read/write/meta latency for each flagged disk. # The low threshold should be higher than 1ms (average seek latency of a 7200 RPM disk is 4ms). Let's conservatively set this to 20ms. # startDiskOutlierDetectionThread should call Thread.currentThread().interrupt() after catching InterruptedException. See https://www.ibm.com/developerworks/library/j-jtp05236/ # slowDiskDetectionThread should be a daemon thread. # DataNodePeerMetrics should also use OutlierDetector constructor that accepts minNumResources, and pass {{10}}, to keep the behavior consistent with what we have. was (Author: arpitagarwal): [~hanishakoneru], a few comments: # {{DataNode#shutdown}} should use slowDiskDetectionThread.join() instead of sleep. # {{diskOutliers}} should maintain the mean read/write/meta latency for each flagged disk. # The low threshold should be higher than 1ms (seek latency of a 7200 RPM disk is 4ms). Let's conservatively set this to 20ms. # startDiskOutlierDetectionThread should call Thread.currentThread().interrupt() after catching InterruptedException. See https://www.ibm.com/developerworks/library/j-jtp05236/ # slowDiskDetectionThread should be a daemon thread. # DataNodePeerMetrics should also use OutlierDetector constructor that accepts minNumResources, and pass {{10}}, to keep the behavior consistent with what we have. > DataNode Disk Outlier Detection > ------------------------------- > > Key: HDFS-11461 > URL: https://issues.apache.org/jira/browse/HDFS-11461 > Project: Hadoop HDFS > Issue Type: Improvement > Components: hdfs > Reporter: Hanisha Koneru > Assignee: Hanisha Koneru > Attachments: HDFS-11461.000.patch > > > Similar to how DataNodes collect peer performance statistics, we can collect > disk performance statistics per datanode and detect outliers among them, if > any. -- This message was sent by Atlassian JIRA (v6.3.15#6346) --------------------------------------------------------------------- To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org