The International Conference for High Performance Computing, Networking, Storage and Analysis
Parallel Complex Network Partitioning.
Student: George M. Slota (Pennsylvania State University)
Supervisor: Kamesh Madduri (Pennsylvania State University)
Abstract: A large number parallel graph analytics follow a bulk synchronous parallel () model: periods of parallel computation followed by periods of parallel communication. To maximize parallel efficiency when a graph is distributed across a cluster, we want a partitioning of the input graph that balances both work and memory (proportional to number of vertices and edges per process) and communication (proportional to total edge cut and maximal edge cut per process). Traditional multi-level partitioners are unable to satisfy all of these requirements and are heavy-weight in terms of computational and memory requirements. This work introduces PULP, an iterative partitioning methodology for small-world graphs that can simultaneously handle multiple constraints and multiple objectives. Partitions produced by PULP are equal to or better than state-of-the-art partitioners in terms of edge cut. PULP also runs very fast, being capable of partitioning multi-billion edge graphs in minutes on a single compute node.