Hello,
I have set up a two-node Hadoop cluster on Ubuntu 12.04 running streaming jobs
with Hadoop 2.2.0. I am having problems with running tasks on a NM which is on
a different host than the RM, and I believe that this is happening because the
NM host's dfs.client.local.interfaces property is not having any effect.
I have two hosts set up as follows:
Host A (1.2.3.4):
NameNode
DataNode
ResourceManager
Job History Server
Host B (5.6.7.8):
DataNode
NodeManager
On each host, hdfs-site.xml was edited to change dfs.client.local.interfaces
from an interface name ("eth0") to the IPv4 address representing that host's
interface ("1.2.3.4" or "5.6.7.8"). This is to prevent the HDFS client from
randomly binding to the IPv6 side of the interface (it randomly swaps between
the IP4 and IP6 addresses, due to the random bind IP selection in the DFS
client) which was causing other problems.
However, I am observing that the Yarn container on the NM appears to inherit
the property from the copy of hdfs-site.xml on the RM, rather than reading it
from the local configuration file. In other words, setting the
dfs.client.local.interfaces property in Host A's configuration file causes the
Yarn containers on Host B to use same value of the property. This causes the
map task to fail, as the container cannot establish a TCP connection to the
HDFS. However, on Host B, other commands that access the HDFS (such as "hadoop
fs") do work, as they respect the local value of the property.
To illustrate with an example, I start a streaming job from the command line on
Host A:
hadoop jar $HADOOP_HOME/share/hadoop/tools/lib/hadoop-streaming-2.2.0.jar
-input hdfs://hosta/linesin/ -output hdfs://hosta/linesout -mapper
/home/hadoop/toRecords.pl -reducer /bin/cat
The NodeManager on Host B notes that there was an error starting the container:
13/12/14 19:38:45 WARN nodemanager.DefaultContainerExecutor: Exception from
container-launch with container ID: container_1387067177654_0002_01_000001 and
exit code: 1
org.apache.hadoop.util.Shell$ExitCodeException:
at org.apache.hadoop.util.Shell.runCommand(Shell.java:464)
at org.apache.hadoop.util.Shell.run(Shell.java:379)
at
org.apache.hadoop.util.Shell$ShellCommandExecutor.execute(Shell.java:589)
at
org.apache.hadoop.yarn.server.nodemanager.DefaultContainerExecutor.launchContainer(DefaultContainerExecutor.java:195)
at
org.apache.hadoop.yarn.server.nodemanager.containermanager.launcher.ContainerLaunch.call(ContainerLaunch.java:283)
at
org.apache.hadoop.yarn.server.nodemanager.containermanager.launcher.ContainerLaunch.call(ContainerLaunch.java:79)
at java.util.concurrent.FutureTask.run(Unknown Source)
at java.util.concurrent.ThreadPoolExecutor.runWorker(Unknown Source)
at java.util.concurrent.ThreadPoolExecutor$Worker.run(Unknown Source)
at java.lang.Thread.run(Unknown Source)
On Host B, I open
userlogs/application_1387067177654_0002/container_1387067177654_0002_01_000001/syslog
and find the following messages (note the DEBUG-level messages which I
manually enabled for the DFS client):
2013-12-14 19:38:32,439 DEBUG [main] org.apache.hadoop.hdfs.DFSClient: Using
local interfaces [1.2.3.4] with addresses [/1.2.3.4:0]
<cut>
2013-12-14 19:38:33,085 DEBUG [main] org.apache.hadoop.hdfs.DFSClient: newInfo
= LocatedBlocks{
fileLength=537
underConstruction=false
blocks=[LocatedBlock{BP-1911846690-1.2.3.4-1386999495143:blk_1073742317_1493;
getBlockSize()=537; corrupt=false; offset=0; locs=[5.6.7.8:50010,
1.2.3.4:50010]}]
lastLocatedBlock=LocatedBlock{BP-1911846690-1.2.3.4-1386999495143:blk_1073742317_1493;
getBlockSize()=537; corrupt=false; offset=0; locs=[5.6.7.8:50010,
1.2.3.4:50010]}
isLastBlockComplete=true}
2013-12-14 19:38:33,088 DEBUG [main] org.apache.hadoop.hdfs.DFSClient:
Connecting to datanode 5.6.7.8:50010
2013-12-14 19:38:33,090 DEBUG [main] org.apache.hadoop.hdfs.DFSClient: Using
local interface /1.2.3.4:0
2013-12-14 19:38:33,095 WARN [main] org.apache.hadoop.hdfs.DFSClient: Failed to
connect to /5.6.7.8:50010 for block, add to deadNodes and continue.
java.net.BindException: Cannot assign requested address
Note the failure to bind to 1.2.3.4, as the IP for Node B's local interface is
actually 5.6.7.8.
Note that when running other HDFS commands on Host B, Host B's setting for
dfs.client.local.interfaces is respected. On host B:
hadoop@nodeb:~$ hadoop fs -ls hdfs://hosta/
13/12/14 19:45:10 DEBUG hdfs.DFSClient: Using local interfaces [5.6.7.8] with
addresses [/5.6.7.8:0]
Found 3 items
drwxr-xr-x - hadoop supergroup 0 2013-12-14 00:40
hdfs://hosta/linesin
drwxr-xr-x - hadoop supergroup 0 2013-12-14 02:01 hdfs://hosta/system
drwx------ - hadoop supergroup 0 2013-12-14 10:31 hdfs://hosta/tmp
If I change dfs.client.local.interfaces on Host A to eth0 (without touching the
setting on Host B), the syslog mentioned above instead shows the following:
2013-12-14 22:32:19,686 DEBUG [main] org.apache.hadoop.hdfs.DFSClient: Using
local interfaces [eth0] with addresses [/<some IP6 address>:0,/5.6.7.8:0]
The job then successfully completes sometimes, but both Host A and Host B will
then randomly alternate between the IP4 and IP6 side of their eth0 interfaces,
which causes other issues. In other words, changing the
dfs.client.local.interfaces setting on Host A to a named adapter caused the
Yarn container on Host B to bind to an identically named adapter.
Any ideas on how I can reconfigure the cluster so every container will try to
bind to its own interface? I successfully worked around this issue by doing a
custom build of HDFS which hardcodes my IP address in the DFSClient, but I am
looking for a better long-term solution.
Thanks,
Jeff
