CAREER: An Adaptive, High-Performance Software Infrastructure for Hierarchical Systems

点击翻译按钮获取中文摘要

Data hierarchies - that is, different access speeds for different areas of computer memory - in parallel computer architectures are becoming increasingly complex. Machines with several levels of memory, including cache-coherent non-uniform memory architectures (CC-NUMA) and clusters of workstations, now dominate supercomputing. In the near future, all high-performance parallel architectures in both science and industry will be hierarchical systems, and the successful development of a general-purpose approach to programming these machines is imperative. Unfortunately, although hierarchical systems are not new, the research community has not yet developed a general and comprehensive approach to them. Fortunately, rapidly decreasing cost of processing power and recent developments in dynamic compilation and runtime support may allow solutions to this problem that were not possible before. This project will build a software infrastructure that exploits those factors to deliver high levels of application performance on hierarchical systems. The infrastructure will support the advance of computational science, serve as a model for future commercial systems, and serve as a tool to educate and expose students to explicit management of parallelism and memory hierarchies. Through collaborations with the National Center for Supercomputing Applications it will also produce a wide range of high-performance scientific and non-scientific codes developed for hierarchical systems.Technically, the project approaches the hierarchy problem with a 4-pronged attack:-A hierarchical virtual machine that abstracts resource allocation and management issues -A hierarchy-aware runtime system that offers the illusion of a non-hierarchical system by adapting an application's communication and synchronization operations-Language constructs and dynamic compiler support to tune application behavior to the hierarchy-Applications that demonstrate the value of the frameworkParts of this approach will receive support from other sources as well as this award.