[jira] [Commented] (HBASE-22306) Improve SLB to distribute table regions across all region servers

["Biju Nair (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HBASE-22306?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16825615#comment-16825615 ] 

Biju Nair commented on HBASE-22306:
-----------------------------------

Attached a modified {{TableSkewCostFunction}} (TableSkewCostFunctionNew.rtf) and quick test results similar to the one documented in the description of the ticket. The tests shows improvement in the distribution but require additional tests. Any comments/suggestions and other alternative approaches?
{noformat}
Cluster: 
Nodes  regions  Tables 
  5      50       5 

Test 1: Initial Distribution
————————— 
Table 0 regions [2, 2, 1, 2, 0]
Table 1 regions [4, 2, 2, 4, 1]
Table 2 regions [2, 2, 3, 1, 0]
Table 3 regions [2, 3, 3, 4, 1]
Table 4 regions [2, 3, 3, 1, 0]

Test 1: After Balancer Run
—————————
Table 0 regions [1, 2, 1, 2, 1]
Table 1 regions [3, 2, 2, 3, 3]
Table 2 regions [2, 1, 2, 1, 2]
Table 3 regions [2, 3, 3, 3, 2]
Table 4 regions [2, 2, 2, 1, 2]

Test 2: Initial Distribution
————————— 
Table 0 regions [3, 0, 3, 2, 1]
Table 1 regions [5, 0, 2, 3, 5]
Table 2 regions [2, 0, 2, 2, 3]
Table 3 regions [2, 1, 1, 0, 1]
Table 4 regions [0, 1, 4, 5, 2]

Test 2: After Balancer Run
—————————
Table 0 regions [2, 2, 2, 2, 1]
Table 1 regions [3, 3, 3, 3, 3]
Table 2 regions [2, 2, 2, 1, 2]
Table 3 regions [1, 1, 1, 1, 1]
Table 4 regions [2, 2, 2, 3, 3]{noformat}

> Improve SLB to distribute table regions across all region servers
> -----------------------------------------------------------------
>
>                 Key: HBASE-22306
>                 URL: https://issues.apache.org/jira/browse/HBASE-22306
>             Project: HBase
>          Issue Type: Improvement
>          Components: Balancer
>            Reporter: Biju Nair
>            Priority: Major
>              Labels: balancer
>         Attachments: TableSkewCostFunctionNew.rtf
>
>
> Noticed in clusters that distribution of regions of tables are skewed. From quick tests, it looks like that the current table skew cost may not be having any influence on selecting the target "cluster" candidate.
> {noformat}
> Cluster: 
> Nodes  regions  Tables 
>   5      50       5 
> Test 1: Initial Distribution
> ———————————
> Table 0 regions [1, 1, 2, 4, 2]
> Table 1 regions [1, 2, 4, 2, 2]
> Table 2 regions [0, 2, 2, 0, 4]
> Table 3 regions [0, 5, 0, 2, 3]
> Table 4 regions [0, 2, 4, 4, 1]
> Test 1: After Balancer Run
> ———————————
> Table 0 regions [2, 2, 2, 2, 2]
> Table 1 regions [2, 2, 3, 2, 2]
> Table 2 regions [1, 2, 2, 2, 1]
> Table 3 regions [2, 3, 0, 2, 3]
> Table 4 regions [3, 1, 3, 2, 2]
> Test 2: Initial Distribution
> ———————————
> Table 0 regions [2, 1, 0, 1, 2]
> Table 1 regions [5, 1, 1, 2, 1]
> Table 2 regions [2, 4, 0, 2, 4]
> Table 3 regions [1, 4, 1, 3, 1]
> Table 4 regions [2, 2, 0, 4, 4]
> Test 2: After Balancer Run
> ———————————-
> Table 0 regions [1, 2, 2, 1, 0]
> Table 1 regions [2, 2, 2, 2, 2]
> Table 2 regions [3, 3, 1, 2, 3]
> Table 3 regions [2, 2, 2, 2, 2]
> Table 4 regions [2, 1, 3, 3, 3]{noformat}



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)