Make EC2 cluster nodes more independent of each other

• Key: HADOOP-2410
• Type: Improvement
• Status: Closed
• Resolution: Fixed
• Priority: Major
• Assignee: Chris K Wensel
• Reporter: Tom White
• Component/s: contrib/ec2
• Affects Version/s: 0.16.1
• Fix Version/s: 0.17.0

Description

The cluster start up scripts currently wait for each node to start up before appointing a master (to run the namenode and jobtracker on), and copying private keys to all the nodes, and writing the private IP address of the master to the hadoop-site.xml file (which is then copied to the slaves via rsync). Only once this is all done is hadoop started on the cluster (from the master). This can fail if any of the nodes fails to come up, which can happen as EC2 doesn't guarantee that you get a cluster of the size you ask for (I've seen this happen).

The process would be more robust if each node was told the address of the master as user metadata and then started its own daemons. This is complicated by the fact that the public DNS alias of the master resolves to a public IP address so cannot be used by EC2 nodes (see http://docs.amazonwebservices.com/AWSEC2/2007-08-29/DeveloperGuide/instance-addressing.html). Instead we need to use a trick (http://developer.amazonwebservices.com/connect/message.jspa?messageID=71126#71126) to find the private IP, and what's more we need to attempt to resolve the private IP in a loop until it is available since the DNS will only be set up after the master has started.

This change will also mean the private key doesn't need to be copied to each node, which can be slow and has dubious security. Configuration can be handled using the mechanism described in HADOOP-2409.

Comments

[Tom White]

Another motivation for this change is to make it more straightforward to add nodes on the fly to an existing cluster. The only change needed would be an extra parameter in the launch script to indicate whether to start a master node - if set to true instance 0 would start as the master (as it does at the moment), otherwise all the new instances would connect to an existing master.

[Tom White]

Amazon has just released a new feature, EC2 Availability Zones (http://developer.amazonwebservices.com/connect/entry.jspa?externalID=1347&categoryID=112), which give some control over instance placement. This would be useful for adding nodes on the fly to an existing cluster to ensure that nodes are in the same zone.

Also, it is now possible to use the public IP address of EC2 nodes from within the EC2 cluster (contrary to the comment in the description above). However, this will incur data transfer costs, which can be avoided by using the private IP address. See http://docs.amazonwebservices.com/AWSEC2/2008-02-01/DeveloperGuide/instance-addressing.html.

[Chris K Wensel]

Here is a first cut at supporting multiple concurrent clusters, the image instance sizes, zone availability, and ganglia.

[Chris K Wensel]

Patch

[Chris K Wensel]

This patch represents a fair number of changes and will need accompanying documentation.

The typical usecase is this:

> hadoop-ec2 launch-cluster my-group 5
> hadoop-ec2 push my-group path/to/some.jar
> hadoop-ec2 login my-group
> hadoop-ec2 terminate-cluster my-group

In another window (after launch-cluster), this is quite useful, and works will with FoxyProxy:
> hadoop-ec2 proxy my-group

There are still some rough edges I think.

> hadoop-ec2
Usage: hadoop-ec2 COMMAND
where COMMAND is one of:
list list all running Hadoop EC2 clusters
launch-cluster <group> <num slaves> launch a cluster of Hadoop EC2 instances - launch-master then launch-slaves
launch-master <group> launch or find a cluster master
launch-slaves <group> <num slaves> launch the cluster slaves
terminate-cluster terminate all Hadoop EC2 instances
login <group|instance id> login to the master node of the Hadoop EC2 cluster
screen <group|instance id> start or attach 'screen' on the master node of the Hadoop EC2 cluster
proxy <group|instance id> start a socks proxy on localhost:6666 (use w/foxyproxy)
push <group> <file> scp a file to the master node of the Hadoop EC2 cluster
<shell cmd> <group|instance id> execute any command remotely on the master
create-image create a Hadoop AMI

[Hadoop QA]

-1 overall. Here are the results of testing the latest attachment
http://issues.apache.org/jira/secure/attachment/12378761/concurrent-clusters.patch
against trunk revision 619744.

@author +1. The patch does not contain any @author tags.

tests included -1. The patch doesn't appear to include any new or modified tests.
Please justify why no tests are needed for this patch.

patch -1. The patch command could not apply the patch.

Console output: http://hudson.zones.apache.org/hudson/job/Hadoop-Patch/2084/console

This message is automatically generated.

[Chris K Wensel]

submitted with correct paths

[Chris K Wensel]

correct path offset

[Chris K Wensel]

removed DFS_WRITE_RETRIES as it's not really necessary with the new kernels.

[Chris K Wensel]

a tar of all relevant files.

[Tom White]

Chris,

These changes look great. Thanks! I'll try them out next week.

[Chris K Wensel]

this version checks both current releases and archives for the hadoop distro

[Tom White]

I've just committed this. Thanks Chris!

I tried out the new scripts and they worked fine. I changed the version of Hadoop in the env file to be 0.17.0 so that it picks up the new AMI when it is created (after 0.17.0 is released). I also changed the version of Java to 1.6.0_05.

Chris, could you update the documentation on the wiki page with the changes please? It would be worth keeping the instructions for the older scripts around on the same page.

[Hudson]

Integrated in Hadoop-trunk #451 (See http://hudson.zones.apache.org/hudson/job/Hadoop-trunk/451/)

[Nate Carlson]

One change I've had to made is to add the memory option for the processes the cluster launches in the hadoop-site.xml that gets generated.. this would probably be a good thing to make configurable for the end user.

I also have manually installed nph-proxy on the master node, with http authentication – makes it much easier to get around the slave nodes.

[Chris K Wensel]

Good idea re configurable memory for the trackers and datanode services, though I find the defaults fine (so far). But I tend to pass my child vm option in per job since they vary. Still a good idea to provide the option.

Note that:

hadoop-ec2 proxy <cluster-name>

starts a local SOCKS tunnel. Used with FoxyProxy FF plugin, you can browse your cluster.