[
https://issues.apache.org/jira/browse/HDFS-4253?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13509310#comment-13509310
]
Colin Patrick McCabe commented on HDFS-4253:
--------------------------------------------
More about that {{compare}} method: I re-read this method, with an eye to
seeing if it satisfied all the comparator requirements outlined in Effective
Java, listed here: http://www.javapractices.com/topic/TopicAction.do?Id=10 I
don't think that it does. In particular, you need to satisfy the "if
a.compare(b) < 0, b.compare(a) > 0" requirement.
I understand now why you special-cased a == b; it's so that the next few lines,
where you test that a == rdr and b == rdr, they can't both be true at that
point. While this is correct, it's also non-obvious, so I would suggest either
adding a comment or simply testing {{if (aIsLocal && bIsLocal)}} directly.
The bug comes later, where you always return 1 if neither Node is on the local
rack. This is wrong; it violates anticommutation (see link). I would suggest
something along these
lines:
{code}
salt = Random.nextInt();
...
@Override
public int compare(Node a, Node b) {
return ComparisonChain.start()
.compare(a != rdr, b != rdr)
.compare(!isOnSameRack(a, rdr), !isOnSameRack(b, rdr))
.compare(a.getHash() ^ salt, b.getHash() ^ salt)
.compare(a.getName(), b.getName());
.result();
}
{code}
> block replica reads get hot-spots due to NetworkTopology#pseudoSortByDistance
> -----------------------------------------------------------------------------
>
> Key: HDFS-4253
> URL: https://issues.apache.org/jira/browse/HDFS-4253
> Project: Hadoop HDFS
> Issue Type: Bug
> Affects Versions: 3.0.0, 2.0.2-alpha
> Reporter: Andy Isaacson
> Assignee: Andy Isaacson
> Attachments: hdfs4253-1.txt, hdfs4253.txt
>
>
> When many nodes (10) read from the same block simultaneously, we get
> asymmetric distribution of read load. This can result in slow block reads
> when one replica is serving most of the readers and the other replicas are
> idle. The busy DN bottlenecks on its network link.
> This is especially visible with large block sizes and high replica counts (I
> reproduced the problem with {{-Ddfs.block.size=4294967296}} and replication
> 5), but the same behavior happens on a small scale with normal-sized blocks
> and replication=3.
> The root of the problem is in {{NetworkTopology#pseudoSortByDistance}} which
> explicitly does not try to spread traffic among replicas in a given rack --
> it only randomizes usage for off-rack replicas.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
