Wednesday, November 30, 2011

NFS exported HDFS (CDH3)


For some reasons it could be a good idea to make a hdfs filesystem available across networks as a exported share. Here I describe a working scenario with linux and hadoop with tools both have on board.
I used fuse and libhdfs to mount a hdfs filesystem. Change namenode.local and <PORT> to fit your environment.

Install:
 yum install hadoop-0.20-fuse.x86_64 hadoop-0.20-libhdfs.x86_64

Create a mountpoint:
 mkdir /hdfs-mount

Mount your hdfs (testing):
 hadoop-fuse-dfs dfs://namenode.local:<PORT> /hdfs-mount -d

You will show like that:
 INFO fuse_options.c:162 Adding FUSE arg /hdfs-mount
 INFO fuse_options.c:110 Ignoring option -d
 unique: 1, opcode: INIT (26), nodeid: 0, insize: 56
 INIT: 7.10
 flags=0x0000000b
 max_readahead=0x00020000
 INFO fuse_init.c:101 Mounting namenode.local:<PORT>
 INIT: 7.8
 flags=0x00000001
 max_readahead=0x00020000
 max_write=0x00020000
 unique: 1, error: 0 (Success), outsize: 40

Hit crtl-C after you see "Success".

Make the mount available at boot time:
 echo "hadoop-fuse-dfs#dfs://namenode.local:<PORT> /hdfs-mount fuse usetrash,rw 0 0" >> /etc/fstab

Test:
#> mount -a
#> mount
 [..]
 sunrpc on /var/lib/nfs/rpc_pipefs type rpc_pipefs (rw)
 fuse on /hdfs-mount type fuse (rw,nosuid,nodev,allow_other,default_permissions)

To tune the memory for each JVM process take a look into /etc/default/hadoop-0.20-fuse and adjust the settings there.

Export via NFS (unsecure):
First we have to decide which user we use, I suppose the user hdfs. Use "id hdfs":
 uid=104(hdfs) gid=105(hdfs) groups=105(hdfs),104(hadoop) context=root:staff_r:staff_t:SystemLow-SystemHigh

Create an exports-file:
 cat /etc/exports
 /hdfs-mount/user    (fsid=111,rw,wdelay,anonuid=104,anongid=105,sync,insecure,no_subtree_check,no_root_squash)

Expl.: read-write, fsid=unused ID (man 5 exports), write-delay, hdfs user, sync

To export only the user-directory from HDFS prevents you from unwanted changes in system relevant directories (mapred as example).
Restart your NFS Server (service nfs restart).

Now you can use your hdfs as a "local" filesystem, which makes some tasks easier. Note that the "use user" are mapped to the local user, to using root is a bad idea.
Mount the exported NFS on your machine and create / copy your jobdefinitions or files simply.

PS: works only from kernel 2.6.27 upwards

Saturday, November 19, 2011

All in one HDFS Cluster for your pocket

Update 1 (Nov 21, 2011):
- added 3rd interface as host-only-adapter (hadoop1)
- enabled trusted device eth2

About one year ago, I created a small XEN-environment for my engineering pourposes. When I was traveling for hours it was very helpful to track some issues or test new features. The problem was that I had to carry 2 notebooks with me. That was the reason I switched to VirtualBox [1] which runs on OSX, Linux and Windows as well. I could play with my servers and when I did, they configured to death and I reimported them into a clean setup. I think that will also be a good start for new people who have to find into the hadoop ecosystem to see the power without the harm of configuration in a multi-node environment.
The appliance is created with VirtualBox, because it runs on OSX and Windows very easily. The idea behind it is to check new settings in a small environment rather easily; the appliance is designed for research, not for development and really not for production. The appliance has 4 nodes, one master and 3 slaves. The setup is not perfect, but it matched the environment I created it for. We have no seperate secondary namenode, for example. I set up hdfs, hive with mysql-metastore, hBase in distributed mode with zookeeper and stargate.

Before we can play with our own LAB we have to consider that we need some specials before. Please read the site [2] I created for.

[1] https://www.virtualbox.org/wiki/Downloads
[2] http://mapredit.blogspot.com/p/all-in-one-hadoop-multi-node-appliance.html

Thursday, November 3, 2011

HDFS debugging scenario


The first step to debug issues in a running hadoop - environment to take a look at the stacktraces, easy accessible over jobtracker/stacks and let you show all running stacks in a jstack view. You will see the running processes, as an example I discuss a lab testing scenario, see below.

http://jobtracker:50030/stacks

Process Thread Dump: 
43 active threads
Thread 3203101 (IPC Client (47) connection to NAMENODE/IP:9000 from hdfs):
  State: TIMED_WAITING
  Blocked count: 6
  Waited count: 7
  Stack:
    java.lang.Object.wait(Native Method)
    org.apache.hadoop.ipc.Client$Connection.waitForWork(Client.java:676)
    org.apache.hadoop.ipc.Client$Connection.run(Client.java:719)

In that case the RPC connection has a state "TIMED_WAIT" in a block and waited count. That means, the namenode could not answer the RPC request fast enough. The problem belongs the setup as I see often in production environments.
For demonstration I use a ESX Cluster with a VM for the namenode. The ESX abstraction layer for networks isn't performant enough and block the requests. It is always a good idea to use physical servers for infrastructure and services.
Another problem I figured out depends on HP Bladecenter switches from Nortel, a newer update set a hidden switch "dos-filter", disable it. The switch will block all traffic which looks like a DOS attack. That's a serious bug and I wondering why such params are delivered and enabled per default.

Now we take a closer look at the namenode:

With "jps" you can list all running java-processes:
jps
24158 SecondaryNameNode
31684 FlumeMaster
7898 JobTracker
18613 NameNode
16631 Jps
31653 FlumeWatchdog

We check the logs with "tail -f /var/log/hadoop-0.20/*.log|grep -i error". If you sure that all things are well you should look at the java-threads on the jobtracker. With "top -Hc" you'll see the threads and the running command:

Mem:   4043792k total,  3916788k used,   127004k free,   352684k buffers
Swap:  5242864k total,  1448628k used,  3794236k free,   653296k cached


  PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND
18448 hdfs      17   0 2345m 102m  12m S 55.1  1.3   0:01.66 /usr/java/jdk1.6.0_23/bin/java -Dproc_jar -Xmx2000m -Dhadoo
18458 hdfs      16   0 2345m 102m  12m S 31.5  1.3   0:00.95 /usr/java/jdk1.6.0_23/bin/java -Dproc_jar -Xmx2000m -Dhadoo
18457 hdfs      15   0 2345m 102m  12m R 30.9  1.3   0:00.93 /usr/java/jdk1.6.0_23/bin/java -Dproc_jar -Xmx2000m -Dhadoo
18356 hdfs      15   0 2510m 544m  11m S  1.3  6.8   0:00.06 jsvc.exec -Dproc_datanode -pidfile /usr/lib/hadoop-0.20/pid
18366 hdfs      18   0 2345m 102m  12m S  0.7  1.3   0:00.02 /usr/java/jdk1.6.0_23/bin/java -Dproc_jar -Xmx2000m -Dhadoo

to check the IO load on a running job you can use vmstat (sysstat-package):

vmstat -n 2 
procs -----------memory---------- ---swap-- -----io---- --system-- -----cpu------
 r  b   swpd   free   buff  cache   si   so    bi    bo   in   cs us sy id wa st
 0  0 1707244 368124 296528 482624    0    0     0   276 1626 1336 14 17 68  1  0
 0  0 1707244 367752 296544 482724    0    0     0   144 1362  716  2  1 97  0  0
 0  0 1707244 366884 296564 482728    0    0     8    66 1268  750  2  3 94  1  0
 1  0 1707244 366216 296588 482736    0    0     8   260 1269  672  2  4 94  1  0

Here we see a snippet with a running hive-job, the system has no problems with IO, but the interrupts and context switching looks a bit high. That depends on setup, it is always a bad idea to run all services on a single server instance.

Ex. Solution:
The namenode is going deeply into swap, so the first case should be to add more RAM (it was a testcluster with 5 nodes and a VM as namenode to reproduce some timing errors).
Remember, each block of HDFS take 4kb RAM usage on the namenode. The best case here will be to split the services in single instances (jobtracker, namenode and secondary namenode should be outsourced).

Another effective way would be to write a debugscript [1] and use setStatus() and incrCounter() methods on Reporter. Let run your debug script in streaming mode and use the cmd-line-options "-mapdebug" and "-reducedebug".

[1] http://hadoop.apache.org/common/docs/current/mapred_tutorial.html#Debugging