[
https://issues.apache.org/jira/browse/HADOOP-8473?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13288441#comment-13288441
]
Junping Du commented on HADOOP-8473:
------------------------------------
Hi Timer,
Thanks for the comments. NodeGroup layer is trying to map to hypervisor
layer in virtualization environment. So if whole cluster is only a single
physical host with a bunch of VMs running on top of, then nothing can help on
data availability. In fact, enhancing data availability is one reason we want
to deliver in this JIRA Umbrella (HADOOP-8468) for hadoop running in cloud
(with a high percentage built on virtualized platform). i.e: We want to let
hadoop can aware "NodeGroup" layer so that not putting two replicas on two VMs
running on the same physical host. Thoughts?
Thanks,
Junping
> Update Balancer to support new NetworkTopology with NodeGroup
> -------------------------------------------------------------
>
> Key: HADOOP-8473
> URL: https://issues.apache.org/jira/browse/HADOOP-8473
> Project: Hadoop Common
> Issue Type: Sub-task
> Components: util
> Affects Versions: 1.0.0, 2.0.0-alpha
> Reporter: Junping Du
> Assignee: Junping Du
>
> Since the Balancer is a Hadoop Tool, it was updated to be directly aware of
> four-layer hierarchy instead of creating an alternative Balancer
> implementation. To accommodate extensibility, a new protected method,
> doChooseNodesForCustomFaultDomain is now called from the existing chooseNodes
> method so that a subclass of the Balancer could customize the balancer
> algotirhm for other failure and locality topologies. An alternative option is
> to encapsulate the algorithm used for the four-layer hierarchy into a
> collaborating strategy class.
> The key changes introduced to support a four-layer hierarchy were to override
> the algorithm of choosing <source, target> pairs for balancing. Unit tests
> were created to test the new algorithm.
> The algorithm now makes sure to choose the target and source node on the same
> node group for balancing as the first priority. Then the overall balancing
> policy is: first doing balancing between nodes within the same nodegroup then
> the same rack and off rack at last. Also, we need to check no duplicated
> replicas live in the same node group after balancing.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
