| Popis: |
Complex systems such as social, biological and information networks, characterized by millions to billions of sub-entities and relationships between them are best represented by graphs. A distinguishing feature of such complex systems is the self-organization into dense clusters called communities. Detecting such communities in massive graphs is critical to the understanding of such complex systems. However community detection is non-trivial and is expensive due to a vast number of computations needed to fully realize the underlying relationships. While a number of near-exact and heuristic algorithms is available for community detection, parallelizing such algorithms to fully leverage the advantages of parallel hardware is still a challenging problem. Most presently available approaches attempt to optimize run-time complexities by scaling original serial community detection algorithms. Graph algorithms in general and existing community detection algorithms in particular are known for not being commensurate with linear scalability on parallel systems. Therefore, there is a need for a combination of high-performance software and a hardware platform that would support efficient parallel graph processing with respect to community detection. We present an Intel® Xeon Phi Label Propagation algorithm (PLPA) variant of community detection algorithm based on label propagation (LP). Our algorithm was tuned for the Intel® Xeon Phi platform a novel architecture that provides 50+ physical cores with simultaneous multithreading. We outline Phi architecture advantages and limitations for PLPA and massive graph processing in general. We present test results of running our algorithm on large real-world networks while achieving near linear speedups and improving the quality of the detected communities. We also analyze possibilities of processing massive networks that cannot fully fit in a Phi memory and hence we extend our initial solution to a modified PLPA (PLPA-M). |
