EECS Seminar: Parallel Graph Processing on GPUs and Clusters

Abstract: The importance of iterative graph algorithms has grown due to their widespread use in graph mining and analytics. Although computations on graphs with millions of nodes and edges contain vast amounts of data level parallelism, exploiting this parallelism is challenging due to the highly irregular nature of real-world graphs. In this talk I will present our recent results that greatly improve the SIMD efficiency and communication efficiency of graph processing on GPUs and clusters. In comparison to prior techniques, our Warp Segmentation technique achieves 1.3x-2.8x performance improvement on a single GPU, our Vertex Refinement technique achieves 2.7x performance improvement on a multi-GPU system, and our Relaxed Consistency protocol achieves 2.3x performance improvement on a 16-node cluster.

Bio: Gupta is a professor of computer science at the University of California, Riverside. His research interests include compilers, architectures and runtimes for parallel systems; and tools for monitoring and debugging parallel software. He has supervised Ph.D. dissertations of 24 students including two winners of the ACM SIGPLAN Outstanding Doctoral Dissertation Award. Papers coauthored by Gupta with his students have been selected for inclusion in 20 Years of PLDI (1979-1999), a best paper award in PACT 2010, and a distinguished paper award in ICSE 2003. Gupta is a Fellow of the ACM, IEEE and AAAS. He received the National Science Foundation's Presidential Young Investigator Award and UCR Doctoral Dissertation Advisor/Mentor Award. He has chaired several major conferences including FCRC, PLDI, HPCA, ASPLOS, CGO, CC, HiPEAC and LCTES. He serves on the editorial boards of ACM Transactions on Architecture and Code Optimization and Parallel Computing journal. Gupta served on a technical advisory group on networking and information technology created by the PCAST.