hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Tao Jie (JIRA)" <j...@apache.org>
Subject [jira] [Assigned] (HDFS-13032) Make AvailableSpaceBlockPlacementPolicy more adaptive
Date Wed, 20 Jun 2018 01:15:00 GMT

     [ https://issues.apache.org/jira/browse/HDFS-13032?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel

Tao Jie reassigned HDFS-13032:

    Assignee: Tao Jie

> Make AvailableSpaceBlockPlacementPolicy more adaptive
> -----------------------------------------------------
>                 Key: HDFS-13032
>                 URL: https://issues.apache.org/jira/browse/HDFS-13032
>             Project: Hadoop HDFS
>          Issue Type: Improvement
>    Affects Versions: 2.8.2
>            Reporter: Tao Jie
>            Assignee: Tao Jie
>            Priority: Major
>         Attachments: HDFS-13032.001.patch, HDFS-13032.002.patch
> In a heterogeneous HDFS cluster, datanode capacity and usage are very different.
> Now we can use HDFS-8131, a usage-aware block placement policy to deal with the problem.
However, this policy could be more flexible.
> 1, The probability of a node with high usage being chosen is fixed once the parameter
is set. That is the probability is always the same no matter its usage is 90% or 70%. When
the usage of a node is close to full, its probability of being chosen should be lower.
> 2, When the difference of usage is below 5%(hard code), the two nodes are considered
the same usage. I think it's OK when usage is 30% and 35%, but when usage is 93% and 98%,
they should not be treated equally. The correction of probability could be more smooth.
> In my opinion, when we choose one node from two candidates (A: usage 30%, B: usage 60%),
we can calculate the probability according to the available storage. p(A) = 70%/(70% + 40%),
p(B) = 40% (70% +40%). When a node is close to full, the probability would be very small.
> Also we could have another factor to weaken this correctness, and make the modification
not so aggressive.
> Any thought? [~liushaohui]

This message was sent by Atlassian JIRA

To unsubscribe, e-mail: hdfs-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-help@hadoop.apache.org

View raw message