You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@trafodion.apache.org by "Qifan Chen (JIRA)" <ji...@apache.org> on 2015/09/25 17:22:04 UTC
[jira] [Commented] (TRAFODION-1500) ESP co-location needs to
consider some nodes may host too many ESP processes due to data in-balance
[ https://issues.apache.org/jira/browse/TRAFODION-1500?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14908140#comment-14908140 ]
Qifan Chen commented on TRAFODION-1500:
---------------------------------------
Smooth the node map to reassign entries with identical node Id of higher
frequency than the rest to some other nodes.
Assume there are m total entries in the map, and
n nodes in the cluster. Assume further there are s entries in the map (out of m) that refer
to very few nodes. These s entrres are the subject of smoothing operation. We do so by
1. allocate an array nodeUsageMap[] and the ith entry in it contains all indexes k in the map
that points at i (i.e., nodeMap.getNodeId[k] = i)
2. Find out s by frequency counting
3. Find out m - s
4. Find out how many nodes in each of the s entries that should be moved
The nodels accepting the reasssignment will be from the set of the nodes contained in the map.
// Example . Assume the original node map with 6 entries as follows:
//
// NodeMapIndex 0 1 2 3 4 5
// NodeMapEntry 0 1 1 3 1 5
//
// The nodes accepting the reasssignment will be { 0, 3, 5 }
//
// The subset of entries referring to a few nodes with high frequency: s = { 1, 2, 4 }
// Since f(0)=f(3)=f(5)=1, the average frequency of nodes not in s: (1+1+1)/3 = 1.
// We will allow one (1) assigment in s to remain in node 1 since it is the average frequency, and
// re-assign the rest (marked X) to different nodes via round-robin starting the 1st node in the map.
// The node (1) is excluded from the re-assignment.
//
// NodeMapIndex 0 1 2 3 4 5
// NodeMapEntry 0 1 1 3 1 5
// X X
// ^ ^
// | |
// to 0 to 3
//
// The final smoothed node map:
//
// NodeMapIndex 0 1 2 3 4 5
// NodeMapEntry 0 1 0 3 3 5
> ESP co-location needs to consider some nodes may host too many ESP processes due to data in-balance
> ---------------------------------------------------------------------------------------------------
>
> Key: TRAFODION-1500
> URL: https://issues.apache.org/jira/browse/TRAFODION-1500
> Project: Apache Trafodion
> Issue Type: Bug
> Components: sql-cmp
> Reporter: Qifan Chen
>
> When a table is range-partitioned, or stats partitioned, it is possible that the amount of the data hosted at each region server may not be uniform. Assume region server s running on node n hosts more data than the rest of the region servers. With the current ESP co-location assignment, node s will be assigned with more # of ESPs.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)