Doctor of Philosophy (PhD)
Large-scale graph applications are of great national, commercial, and societal importance, with direct use in ﬁelds such as counter-intelligence, proteomics, and data mining. Unfortunately, graph-based problems exhibit certain basic characteristics that make them a poor match for conventional computing systems in terms of structure, scale, and semantics. Graph processing kernels emphasize sparse data structures and computations with irregular memory access patterns that destroy the temporal and spatial locality upon which modern processors rely for performance. Furthermore, applications in this area utilize large data sets, and have been shown to be more data intensive than typical ﬂoating-point applications, two properties that lead to inefficient utilization of the hierarchical memory system. Current approaches to processing large graph data sets leverage traditional HPC systems and programming models, for shared memory and message-passing computation, and are thus limited in efficiency, scalability, and programmability. The research presented in this thesis investigates the potential of a new model of execution that is hypothesized as a promising alternative for graph-based applications to conventional practices. A new approach to graph processing is developed and presented in this thesis. The application of the experimental ParalleX execution model to graph processing balances continuation-migration style ﬁne-grain concurrency with constraint-based synchronization through embedded futures. A collection of parallel graph application kernels provide experiment control drivers for analysis and evaluation of this innovative strategy. Finally, an experimental software library for scalable graph processing, the ParalleX Graph Library, is deﬁned using the HPX runtime system, providing an implementation of the key concepts and a framework for development of ParalleX-based graph applications.
Document Availability at the Time of Submission
Release the entire work immediately for access worldwide.
Stark, Dylan Thomas, "Advanced semantics for accelerated graph processing" (2011). LSU Doctoral Dissertations. 3131.