[
https://issues.apache.org/jira/browse/HDFS-7300?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14187195#comment-14187195
]
Kihwal Lee commented on HDFS-7300:
----------------------------------
The base formula for caluculating the value is
{noformat}
maxNodesPerRack = (totalNumOfReplicas-1)/numOfRacks + 2
{noformat}
In the patch, the single rack case and the single replica case are handled
without applying this formula. Then it is guaranteed that the number of rack is
greater than 1 when calculating the max value. It is also guaranteed to give a
sufficiently big max value.
{noformat}
maxNodePerRack * numOfRacks >= totalNumOfReplicas
totalNumOfReplicas-1 + 2 * numOfRack >= totalNumOfReplicas
numOfRack >= 0.5
{noformat}
Since numOfRacks is greater than 1, maxNodePerRack is guaranteed to be large
enough.
In order to take care of the case of {{maxNodePerRack == totalNumOfReplicas}},
which happens in the cases described in the description, maxNodePerRack is
decremented if necessary. This still results in a sufficiently large value.
{noformat}
(maxNodePerRack - 1) * numOfRacks > totalNumOfReplicas
totalNumOfReplicas-1 + numOfRack > totalNumOfReplicas
numOfRack > 1
{noformat}
It shows the resulting max value is not only large enough, but also allows a
bit of slack for unbalanced racks, as the original formula does.
> The getMaxNodesPerRack() method in BlockPlacementPolicyDefault is flawed
> ------------------------------------------------------------------------
>
> Key: HDFS-7300
> URL: https://issues.apache.org/jira/browse/HDFS-7300
> Project: Hadoop HDFS
> Issue Type: Bug
> Reporter: Kihwal Lee
> Assignee: Kihwal Lee
> Priority: Critical
> Attachments: HDFS-7300.patch
>
>
> The {{getMaxNodesPerRack()}} can produce an undesirable result in some cases.
> - Three replicas on two racks. The max is 3, so everything can go to one rack.
> - Two replicas on two or more racks. The max is 2, both replicas can end up
> in the same rack.
> {{BlockManager#isNeededReplication()}} fixes this after block/file is closed
> because {{blockHasEnoughRacks()}} will return fail. This is not only extra
> work, but also can break the favored nodes feature.
> When there are two racks and two favored nodes are specified in the same
> rack, NN may allocate the third replica on a node in the same rack, because
> {{maxNodesPerRack}} is 3. When closing the file, NN moves a block to the
> other rack. There is 66% chance that a favored node is moved. If
> {{maxNodesPerRack}} was 2, this would not happen.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
