CAREER: Programming the Existing and Emerging Memory Systems for Extreme-scale Parallel Performance

点击翻译按钮获取中文摘要

High performance computing (HPC) focuses on using numerical model to simulate complex science and engineering phenomena, such as galaxies, weather and climate, molecular interactions, electric power grids, and aircraft in flight. Over the next decade the goal is to build HPC parallel system capable of extreme-scale performance (one exaflop (1018)operations per second) and processing exabyte (1018) of data. However, one of the biggest challenges of achieving extreme-scale performance is what is known as the hardware memory wall, which is about the growing gap between the speed of computation performed by CPU and the speed of supplying data to the CPU from memory systems (about x100 time slower). The low performance efficiency of modern HPC system (average 60% and could be as low as 5%) manifests the memory wall impact since a huge amount of computation cycles are wasted for waiting for the arrival of input data. It becomes very critical to create effective software solutions for achieving the computation potential of hardware and for improving the efficiency and usability of the existing and future computing system. Such solutions will significantly benefit a broad range of disciplines that use parallel computers to solve scientific and engineering problems, and accelerate scientific discovery and problem solving to improve quality of life of the society. This CAREER project develops innovative software techniques to address the programming and performance challenges of the existing and emerging memory systems: 1) a portable abstract machine model for programming, compiling and executing parallel applications, 2) new programming interface and model for data mapping, movement, and consistency, and 3) machine-aware compilation and data-aware scheduling techniques to realize an asynchronous task flow execution model to hide the latency of data movement. It addresses the memory wall challenge by developing a memory-centric programming paradigm for helping achieve extreme-scale performance of parallel applications with minimum impairment to programmability. For education, the project involves a broader community starting from high school in the area of HPC and computer science.