Page History

This document assume that you have already installed Hama cluster and you have tested it using some examples.

PageRank

• Uses the PageRank algorithm described in the Google Pregel paper
• Introduces partitioning and collective communication
  

...

Run PageRank on Hama Cluster

First of all, generate a symmetric adjacency matrix using the gen command.

  % bin/hama jar ../hama-examples-0.x.0-examples.jar pagerankgen <inputsymmetric path>100 <output10 path> [damping factor] [epsilon error] [tasks]

The default parameters for pagerank are:

0.85 0.001

As you can see 0.85 is the damping factor, that is the probability which a user will "randomly" jump to other sides. See the Random Surfer Model.

0.001 is the convergence error, the error will always be measured after an iteration. It tells how much the pagerank of all sites has changed. If you are setting this to a lower factor, it will take more iterations.

Submit your own Web-graph

You can transform your graph as a adjacency list to fit into the input which Hama is going to parse and calculate the Pagerank.

The file that Hama can successfully parse is a TextFile that has the following layout:

Site1\tSite2\tSite3
Site2\tSite3
Site3

This piece of text will adjacent Site1 to Site2 and Site3, Site2 to Site3 and Site3 is a dangling node. As you can see a site is always on the leftmost side (we call it the key-site), and the outlinks are seperated by tabs (\t) as the following elements.

Make sure that every site's outlink can somewhere be found in the file as a key-site. Otherwise it will result in weird NullPointerExceptions.

Then you can run pagerank on it with:

randomgraph 2

This will create a graph with 100 nodes and 1K edges and store 2 partitions on HDFS as the sequence file. You can adjust partition and tasks numbers to fit your cluster. Then, run PageRank using:

  % 

bin/hama jar ../hama-examples-0.x.0-examples.jar pagerank /tmp/input/input.txt /tmp/pagerank-output

Note that based on what you have configured, the paths may be in HDFS or on local disk.

Output

The output is a double value that is between zero and 1.0. Where 1.0 is a very "famous" site.

All pages' rank should sum up to 1.0, otherwise the algorithm is broken.

Implementation

For detailed questions in terms of implementation have a look at my blog. It describes the algorithm and focuses on the main ideas showing implementation things. It contains ancient code from before Hama 0.5 where we introduced the graph API.

randomgraph pagerankresult 4

Submit your own graph

See WriteHamaGraphFilehttp://codingwiththomas.blogspot.com/2011/04/pagerank-with-apache-hama.html