[jira] Commented: (HADOOP-5638) More improvement on block placement performance

["Hairong Kuang (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HADOOP-5638?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12696830#action_12696830 ] 

Hairong Kuang commented on HADOOP-5638:
---------------------------------------

I did the same experiment described in HADOOP-5603. On a full cluster with 3150 nodes, the patch reduced the amount of time on placing a block on two datanodes from 2.1s to less than 0.1s. 

> More improvement on block placement performance
> -----------------------------------------------
>
>                 Key: HADOOP-5638
>                 URL: https://issues.apache.org/jira/browse/HADOOP-5638
>             Project: Hadoop Core
>          Issue Type: Improvement
>          Components: dfs
>            Reporter: Hairong Kuang
>            Assignee: Hairong Kuang
>             Fix For: 0.21.0
>
>         Attachments: excludedList.patch
>
>
> Block placement algorithm currently has an excluded node list, which contains all datanodes that have been visited. This list is implemented as an array list, whose cost of inserting is O(1) but the cost of query "contains" is O( n ), where n is the number of datanodes. This makes the cost of block placement to be O(n*n) when a cluster is full.
> I propose to change the data structure of the excluded node list as a HashMap. So in average, the cost of insertion is O(1) and the cost of query is O(1). This makes the block placement algorithm to be O( n ) in average.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.