CAREER: Autotuning for multicore and manycore architectures: an enhanced feedback-driven approach

对于应用程序开发人员来说，在复杂的体系结构上实现高比例的峰值性能一直是一个长期的挑战。多核处理器和加速器的出现极大地加剧了这个问题。随着每个套接字的内核数量的增加，共享和分布式缓存的深度层次结构，以及即将出现的百亿亿级计算，多核平台给软件开发和应用程序调优带来了前所未有的挑战。本研究面对多核和多核软件开发的挑战，通过改进反馈诊断来改进自动性能调优。通过增强对问题域、程序特征和体系结构特征的了解，可以实现自动调优效率。为此，我们正在开发一套工具，允许规范、收集和综合调优相关信息。然后通过新颖的机器学习模型利用这些丰富的信息，在不同的架构上提供可扩展、可移植和可持续的性能。根据这项研究，我们为本科生建立了一个结构化的跨学科项目，该项目将来自不同STEM学科的教师和学生聚集在一起，探索跨学科的研究问题，其中计算思维、建模和仿真发挥核心作用。在过去的研究中开发的技术将通过节省高端系统的能源和计算周期而产生直接的经济影响。应用程序效率的提高将使研究人员能够执行大规模模拟，这将使包括医学、高能物理、气候建模和纳米技术在内的广泛学科的复杂现象建模成为可能。

Achieving a high fraction of peak performance on complex architectures has been a perennial challenge for application developers. The emergence of multicore processors and accelerators has greatly exacerbated this problem. With an increasing number of cores per socket, deep hierarchies of shared and distributed caches, and exascale computing on the horizon, multicore platforms pose unprecedented challenges for software development and application tuning. This research confronts the challenge of multicore and manycore software development by improving automatic performance tuning through improved feedback diagnostics. Autotuning efficiency is achieved through enhanced knowledge of the problem domain, program features and architectural characteristics. To this end, we are developing a set of tools that allow specification, collection and synthesis of tuning related information. This rich set of information is then exploited through novel machine learning models to deliver scalable, portable and sustainable performance on diverse architectures. In line with this research, we have established, SISTEM, a structured interdisciplinary program for undergraduates that brings together faculty and students from different STEM disciplines to explore cross-cutting research problems in which computational thinking, modeling and simulation play a central role. The techniques developed as past of this research will have a direct economic impact through saved energy and computation cycles on high-end systems. The improved efficiency of applications will allow researchers to execute large-scale simulations that will enable modeling of complex phenomenon in a broad range of disciplines including medicine, high-energy physics, climate modeling and nanotechnology.