Practical Parallel External Memory Algorithms via Simulation of Parallel Algorithms

Abstract: This thesis introduces PEMS2, an improvement to PEMS (Parallel External Memory System). PEMS executes Bulk-Synchronous Parallel (BSP) algorithms in an External Memory (EM) context, enabling computation with very large data sets which exceed the size of main memory. Many parallel algorithms have been designed and implemented for Bulk-Synchronous Parallel models of computation. Such algorithms generally assume that the entire data set is stored in main memory at once. PEMS overcomes this limitation without requiring any modification to the algorithm by using disk space as memory for additional "virtual processors". Previous work has shown this to be a promising approach which scales well as computational resources (i.e. processors and disks) are added. However, the technique incurs significant overhead when compared with purpose-built EM algorithms. PEMS2 introduces refinements to the simulation process intended to reduce this overhead as well as the amount of disk space required to run the simulation. New functionality is also introduced, including asynchronous I/O and support for multi-core processors. Experimental results show that these changes significantly improve the runtime of the simulation. PEMS2 narrows the performance gap between simulated BSP algorithms and their hand-crafted EM counterparts, providing a practical system for using BSP algorithms with data sets which exceed the size of RAM.