Run Hadoop sort benchmark on Amazon EC2

• Key: HADOOP-4382
• Type: Test
• Status: Open
• Priority: Major
• Assignee: Tom White
• Reporter: Tom White
• Component/s: contrib/ec2

Description

By running a benchmark on EC2 we can see how well Hadoop performs, how to tune it, and how performance changes between releases.

Comments

[Tom White]

A script that:

1. Launches a cluster on EC2
2. Waits for the cluster and Hadoop daemons to start
3. Runs a small sort job to warm up the cluster
4. Runs a sort job and emits the job duration
5. Terminates the cluster

Running on an 8 node cluster it took 2742 seconds to sort 32GB of data using the default hadoop-site.xml that the EC2 scripts use. This could be improved by using better settings.

There are several improvements that could be made to the script, in particular in detecting when the cluster is ready to go (the current script waits until 90% of the nodes are up then waits 1 minute for Hadoop to start). There are more ideas here: http://www.nabble.com/Auto-shutdown-for-EC2-clusters-td20132561.html It would also be good to do multiple runs, discard the first and compute an average.

This should be a good basis for running a regular EC2 benchmark from Hudson.

Comments welcome.

[Tom White]

I should say that the 8 node cluster used large EC2 instances (and the namenode/jobtracker is not included in the 8 nodes).

[Nigel Daley]

Looks good Tom. A couple comments:

• should we also run sortvalidation to ensure the sort actually worked?
• what bin dir are you putting the script in?
• perhaps name the script sort-benchmark
• add a line to echo the # minutes into a file as follows for Hudson plot:

  sort_minutes=`expr $

  Unknown macro: {sort_duration}

  / 60`
  echo "YVALUE=$

  Unknown macro: {sort_minutes}

  " > sort_minutes.properties

Argh, Jira wiki notation ate my code snippet.

sort_minutes=`expr ${sort_duration} / 60`
echo "YVALUE=${sort_minutes}" > sort_minutes.properties

[Tom White]

Thanks for the comments Nigel.

New patch incorporating the suggestions. (I've created the patch from the base of Hadoop this time, so the script goes in src/contrib/ec2/bin.)

[Nigel Daley]

+1