ITR: Algorithms for Irregular Discrete Computations on SMPs

点击翻译按钮获取中文摘要

Parallel computing has promised very high performance, but has delivered it only in a narrow range of applications. Exploiting parallelism at the level of large distributed-memory systems is hampered by the cost of message-passing, while shared-memory systems remain mostly small-scale. With the advent of symmetric multiprocessors (SMPs), however, shared-memory on a modest scale is becoming the standard for desktop scientific and engineering applications. Yet very little has been done yet to support effective parallel computing on an SMP beyond basic linear algebra and fast Fourier transforms. Fortunately, preliminary work by the investigators indicates that it is possible to design and implement algorithms for irregular (i.e. graph-based) computations that provide efficient, scalable performance on SMPs. This project will further develop, implement, assess, and refine those algorithms.Technically, the project will focus on science-driven problems that are graph- or geometry-based (e.g. phylogeny tees and watershed modeling) and thus involve irregular computation. It will provide three main benefits for these problems. First, it will develop new algorithms and a library of basic routines to support graph and computational geometry algorithms, leveraging insights from nearly twenty years of theoretical work on Parallel Random Access Memory (PRAM). Second, it will produce an assessment methodology for SMP-based computations using both generated test instances and real-world data, extending recent developments in algorithmic engineering and experimental algorithmics. Finally, it will provide a practical demonstration of high performance computing on desktop SMPs for scientific problems.