Software Systems for Irregular Application on Scalable Multiprocessors

点击翻译按钮获取中文摘要

This project will implement systems software to support irregular parallel programs on scalable multiprocessors. Machines with physically distributed memory offer the best opportunity for scalable performance, but current programming techniques for these machines are cumbersome and often inefficient. This is particularly true for applications with irregular data structures or irregular patterns of communication and control. Two issues of primary concern are data layout and scheduling. Data layout is critical to avoiding communication overhead, but good layout strategies are difficult to determine during compilation because they depend on global properties of the program and on input data. Similarly, scheduling for irregular problems must adapt to varying loads during program execution. For both problems, there is a trade-off between the inflexibility of decisions made during compilation and the overhead of decisions made at runtime. High level language mechanisms to allow the programmer to specify data layout and scheduling strategies, and implement them using a combination of compiler and runtime support will be developed. The emphasis will be on irregular applications: the layouts, for example, will describe the placement of subgraphs of a logically global graph, and the scheduling primitives will allow for both control and data parallelism. Throughout the project's duration applications to motivate and test the systems work will be developed.