[
https://issues.apache.org/jira/browse/YARN-1226?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Kaibo Zhou updated YARN-1226:
-----------------------------
Description:
When I run a mapreduce job which use TableInputFormat to scan a hbase table on
yarn cluser with 140+ nodes, I consistently get very low data locality around
0~10%.
The scheduler is Capacity Scheduler. Hbase and hadoop are integrated in the
cluster with NodeManager, DataNode and HRegionServer run on the same node.
The reason of low data locality is: most machines in the cluster uses IPV6, few
machines use IPV4. NodeManager use "InetAddress.getLocalHost().getHostName()
" to get the host name, but the return result of this function depends on IPV4
or IPV6, see [InetAddress.getLocalHost().getHostName() returns
FQDN](http://bugs.sun.com/view_bug.do?bug_id=7166687).
On machines with ipv4, NodeManager get hostName as:
search042097.sqa.cm4.site.net
But on machines with ipv6, NodeManager get hostName as: search042097.sqa.cm4
if run with IPv6 disabled, -Djava.net.preferIPv4Stack=true, then returns
search042097.sqa.cm4.site.net.
For the mapred job which scan hbase table, the InputSplit contains node
locations of FQDN, e.g. search042097.sqa.cm4.site.net. Because in hbase, the
RegionServers' hostnames are allocated by HMaster. HMaster communicate with
RegionServers and get the region server's host name use java NIO:
clientChannel.socket().getInetAddress().getHostName().
Also see the startup log of region server:
13:06:21,200 INFO org.apache.hadoop.hbase.regionserver.HRegionServer: Master
passed us hostname to use. Was=search042024.sqa.cm4,
Now=search042024.sqa.cm4.site.net
As you can see, most machines in the Yarn cluster with IPV6 get the short
hostname, but hbase always get the full hostname, so the Host cannot matched
(see RMContainerAllocator::assignToMap).This can lead to poor locality.
After I use java.net.preferIPv4Stack to force IPv4 in yarn, I get 70+% data
locality in the cluster.
Thanks,
Kaibo
was:
When I run a mapreduce job which use TableInputFormat to scan a hbase table on
yarn cluser with 140+ nodes, I consistently get very low data locality around
0~10%.
The scheduler is Capacity Scheduler. Hbase and hadoop are integrated in the
cluster with NodeManager, DataNode and HRegionServer run on the same node.
The reason of low data locality is: most machines in the cluster uses IPV6, few
machines use IPV4. NodeManager use "InetAddress.getLocalHost().getHostName()
" to get the host name, but the return result of this function depends on IPV4
or IPV6, see "http://bugs.sun.com/view_bug.do?bug_id=7166687";.
On machines with ipv4, NodeManager get hostName as:
search042097.sqa.cm4.site.net
But on machines with ipv6, NodeManager get hostName as: search042097.sqa.cm4
if run with IPv6 disabled, -Djava.net.preferIPv4Stack=true, then returns
search042097.sqa.cm4.site.net.
For the mapred job which scan hbase table, the InputSplit contains node
locations of FQDN, e.g. search042097.sqa.cm4.site.net. Because in hbase, the
RegionServers' hostnames are allocated by HMaster. HMaster communicate with
RegionServers and get the region server's host name use java NIO:
clientChannel.socket().getInetAddress().getHostName().
Also see the startup log of region server:
13:06:21,200 INFO org.apache.hadoop.hbase.regionserver.HRegionServer: Master
passed us hostname to use. Was=search042024.sqa.cm4,
Now=search042024.sqa.cm4.site.net
As you can see, most machines in the Yarn cluster with IPV6 get the short
hostname, but hbase always get the full hostname, so the Host cannot matched
(see RMContainerAllocator::assignToMap).This can lead to poor locality.
After I use java.net.preferIPv4Stack to force IPv4 in yarn, I get 70+% data
locality in the cluster.
Thanks,
Kaibo
> ipv4 and ipv6 affect job data locality
> --------------------------------------
>
> Key: YARN-1226
> URL: https://issues.apache.org/jira/browse/YARN-1226
> Project: Hadoop YARN
> Issue Type: Improvement
> Components: capacityscheduler
> Affects Versions: 0.23.3, 2.0.0-alpha, 2.1.0-beta
> Reporter: Kaibo Zhou
> Priority: Minor
>
> When I run a mapreduce job which use TableInputFormat to scan a hbase table
> on yarn cluser with 140+ nodes, I consistently get very low data locality
> around 0~10%.
> The scheduler is Capacity Scheduler. Hbase and hadoop are integrated in the
> cluster with NodeManager, DataNode and HRegionServer run on the same node.
> The reason of low data locality is: most machines in the cluster uses IPV6,
> few machines use IPV4. NodeManager use
> "InetAddress.getLocalHost().getHostName()
> " to get the host name, but the return result of this function depends on
> IPV4 or IPV6, see [InetAddress.getLocalHost().getHostName() returns
> FQDN](http://bugs.sun.com/view_bug.do?bug_id=7166687).
> On machines with ipv4, NodeManager get hostName as:
> search042097.sqa.cm4.site.net
> But on machines with ipv6, NodeManager get hostName as: search042097.sqa.cm4
> if run with IPv6 disabled, -Djava.net.preferIPv4Stack=true, then returns
> search042097.sqa.cm4.site.net.
> For the mapred job which scan hbase table, the InputSplit contains node
> locations of FQDN, e.g. search042097.sqa.cm4.site.net. Because in hbase, the
> RegionServers' hostnames are allocated by HMaster. HMaster communicate with
> RegionServers and get the region server's host name use java NIO:
> clientChannel.socket().getInetAddress().getHostName().
> Also see the startup log of region server:
> 13:06:21,200 INFO org.apache.hadoop.hbase.regionserver.HRegionServer: Master
> passed us hostname to use. Was=search042024.sqa.cm4,
> Now=search042024.sqa.cm4.site.net
> As you can see, most machines in the Yarn cluster with IPV6 get the short
> hostname, but hbase always get the full hostname, so the Host cannot matched
> (see RMContainerAllocator::assignToMap).This can lead to poor locality.
> After I use java.net.preferIPv4Stack to force IPv4 in yarn, I get 70+% data
> locality in the cluster.
> Thanks,
> Kaibo
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
