hadoop-mapreduce-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Sandy Ryza (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (MAPREDUCE-5611) CombineFileInputFormat only requests a single location per split when more could be optimal
Date Mon, 06 Jan 2014 21:33:51 GMT

    [ https://issues.apache.org/jira/browse/MAPREDUCE-5611?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13863446#comment-13863446
] 

Sandy Ryza commented on MAPREDUCE-5611:
---------------------------------------

bq. What if, JT is not able to schedule tasks on this node (slot limitation etc). Then it
will pick any random node and schedule the task (having all the blocks non local).
That's right.  However, picking a node with a small fraction of the input data is not much
better than picking a node without any of the input data.  It is only useful to place a task
on a node if the majority of the data is on that node.  There may be more optimal approaches
to this that take into account the number of bytes on each node, but I think using the intersection
is a good start that we know will not cause perf regressions.

bq. What if there is no intersection i.e common nodes for blocks in a split?
The change was proposed to affect the code where we are building splits out of the nodeToBlocks
map.  In this part of the split creation process, there will always be an intersection because
the blocks are all chosen from a specific node.

> CombineFileInputFormat only requests a single location per split when more could be optimal
> -------------------------------------------------------------------------------------------
>
>                 Key: MAPREDUCE-5611
>                 URL: https://issues.apache.org/jira/browse/MAPREDUCE-5611
>             Project: Hadoop Map/Reduce
>          Issue Type: Bug
>    Affects Versions: 1.2.1
>            Reporter: Chandra Prakash Bhagtani
>            Assignee: Chandra Prakash Bhagtani
>         Attachments: CombineFileInputFormat-trunk.patch
>
>
> I have come across an issue with CombineFileInputFormat. Actually I ran a hive query
on approx 1.2 GB data with CombineHiveInputFormat which internally uses CombineFileInputFormat.
My cluster size is 9 datanodes and max.split.size is 256 MB
> When I ran this query with replication factor 9, hive consistently creates all 6 rack-local
tasks and with replication factor 3 it creates 5 rack-local and 1 data local tasks. 
>  When replication factor is 9 (equal to cluster size), all the tasks should be data-local
as each datanode contains all the replicas of the input data, but that is not happening i.e
all the tasks are rack-local. 
> When I dug into CombineFileInputFormat.java code in getMoreSplits method, I found the
issue with the following snippet (specially in case of higher replication factor)
> {code:title=CombineFileInputFormat.java|borderStyle=solid}
> for (Iterator<Map.Entry<String,
>          List<OneBlockInfo>>> iter = nodeToBlocks.entrySet().iterator();
>          iter.hasNext();) {
>        Map.Entry<String, List<OneBlockInfo>> one = iter.next();
>       nodes.add(one.getKey());
>       List<OneBlockInfo> blocksInNode = one.getValue();
>       // for each block, copy it into validBlocks. Delete it from
>       // blockToNodes so that the same block does not appear in
>       // two different splits.
>       for (OneBlockInfo oneblock : blocksInNode) {
>         if (blockToNodes.containsKey(oneblock)) {
>           validBlocks.add(oneblock);
>           blockToNodes.remove(oneblock);
>           curSplitSize += oneblock.length;
>           // if the accumulated split size exceeds the maximum, then
>           // create this split.
>           if (maxSize != 0 && curSplitSize >= maxSize) {
>             // create an input split and add it to the splits array
>             addCreatedSplit(splits, nodes, validBlocks);
>             curSplitSize = 0;
>             validBlocks.clear();
>           }
>         }
>       }
> {code}
> First node in the map nodeToBlocks has all the replicas of input file, so the above code
creates 6 splits all with only one location. Now if JT doesn't schedule these tasks on that
node, all the tasks will be rack-local, even though all the other datanodes have all the other
replicas.



--
This message was sent by Atlassian JIRA
(v6.1.5#6160)

Mime
View raw message