An Object-Oriented Library for Shared-Memory Parallel Simulations

Programming shared-memory multiprocessor systems is becoming increasingly difficult as the gap between memory speed and processor speed increases. At the same time, this class of computer—based on standard microprocessors—is becoming increasingly common as an alternative to traditional mainframes and supercomputers. Programs that are not sympathetic to caches can perform poorly on such systems. Problems include false sharing (unrelated data in a cache block, resulting in coherence misses), inadequate blocking (processing data as often as possible before it moves out of the cache) and poor exploitation of prefetch (fetching data before it is needed). This research addresses the problem of accommodating changes in memory hierarchy cost and cache characteristics through an object-oriented C++ library called OOSH (Object-Oriented Library for Shared Memory). OOSH includes low-level allocators which pad and align objects to cache block boundaries, and primitives for launching processes, locking and synchronization. OOSH also provides support for object blocking—an object-oriented version of blocking, a technique traditionally used in matrix operations. Blocking divides an algorithm’s data into blocks or tiles, which are processed as far as possible before moving on to other data, with the objective of reducing cache misses. Performance of OOSH is evaluated by running selected applications from the Stanford SPLASH benchmarks on the Mint MIPS architecture simulator. Simulations are run both with parameters similar to current multiprocessor architectures, and with higher memory hierarchy costs, to investigate the impact of technology trends. OOSH versions of the applications are compared with the SPLASH implementations, based on Argonne National Laboratories Parmacs macros. The applications are MP3D (particle-based wind tunnel simulation), Barnes-Hut (nbody gravitation simulation) and Water (water molecule simulation). MP3D only needs local synchronization, and uses distributed synchronization: barriers are replaced by