liu-cheng / pagerank

Add a vertex based page rank computing model

Clusteing the vertices in a graph for pagerank

This is a verification of a simple graph clustering idea. Basically, we want to cluster some of the vertices in a graph as super vertices and try to change the computing method accordingly for the super vertices. Here are a few potential benefits of the clustering:

1. It reduces function call of the page rank's vertex computing function assuming a vertex based graph processing framework.
2. There is potential data reuse within the clustered pagerank computing function when combined.
3. With compilation support, the clustering may be done transparently to the users. Eventually, pagerank may be optimized automatically.

At the same time, we also tried to optimize the data layout on top of the clustering. To that end, we define a similarity meric for each vertex pair based on the overlapping of their incoming neighbors. Then we reoranize the graph layout based on the similarity information. According to the experiments, clusetring has negligible influence on the computing time while the data layout reorganization is relativelly beneficial.