hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Colin Patrick McCabe (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HDFS-97) DFS should detect slow links(nodes) and avoid them
Date Mon, 21 Jul 2014 20:04:40 GMT

    [ https://issues.apache.org/jira/browse/HDFS-97?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14069192#comment-14069192
] 

Colin Patrick McCabe commented on HDFS-97:
------------------------------------------

This is also related to hedged reads.  Even if we blacklist a datanode after a timeout / latency
spike has happened, the damage is done.  Hedged reads can avoid the latency spike in the first
place.

> DFS should detect slow links(nodes) and avoid them
> --------------------------------------------------
>
>                 Key: HDFS-97
>                 URL: https://issues.apache.org/jira/browse/HDFS-97
>             Project: Hadoop HDFS
>          Issue Type: Bug
>            Reporter: Runping Qi
>
> The current DFS does not detect slow links (nodes).
> Thus, when a node or its network link is slow, it may affect the overall system performance
significantly.
> Specifically, when a map job needs to read data from such a node, it may progress 10X
slower.
> And when a DFS data node pipeline consists of such a node, the write performance degrades
significantly.
> This may lead to some long tails for map/reduce jobs. We have experienced such behaviors
quite often.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

Mime
View raw message