Transitive closure: an experimental case study of three multithreaded database algorithms on a shared memory multiprocessor

摘要

An experimental performance evaluation of three multithreaded database transitive closure algorithms on a shared memory multiprocessor is described. The algorithms are differentiated on the basis of the dataflow patterns among the threads, namely no dataflow, pipelined dataflow, and network dataflow. Close to linear speedup is achieved by the no dataflow algorithm when there is sufficiently large workload, and sufficient primary memory for processing. Even when communication is through shared memory, the no dataflow algorithm performs significantly better than the algorithms with dataflow. When access to each package of memory is shared by multiple threads, the number of threads significantly affects the performance of the algorithms in limited primary memory. When each page is exclusively accessed by a single thread, performance remains constant when the number of threads is varied.