CAREER:An Introspective Architecture for Manycore Performance and Power Debugging

职业：多核性能和功耗调试的内省架构

• 批准号: 1149557
• 负责人: Guru Prasadh Venkataramani
• 金额: $ 40.02万
• 依托单位: George Washington University
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2012
• 资助国家: 美国
• 起止时间: 2012-03-15 至 2018-02-28
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1149557&HistoricalAwards=false
• 关键词: CAREER; Introspective; Architecture; Manycore; Performance

Rapid advances in transistor technology combined with innovations in computer architecture have led to manycore processors, which can deliver significantly higher software performance. Getting scalable performance (for example, eight times speedup in program execution time on an eight-core processor) is a challenging task even for experienced software developers. This is because parallel programs can suffer from performance problems that either do not exist in sequential applications or that are aggravated by parallel execution. Also, power consumption of manycore applications is a crucial factor limiting their widespread adoption. Unfortunately, analyzing manycore applications for performance and power using just software tools presents challenges due to programs written in multiple languages and source files, libraries in binary-only form, and so on. An efficient hardware-software framework, that can offer insight into performance and power profiles of software execution and subsequently facilitate remediation mechanisms to overcome the observed program bottlenecks, would be a valuable step toward realizing the promise of manycore computing. This research project explores a transformative hardware-software cooperative solution, where the software receives an introspective look into the hardware behavior and dynamically deploys remediation mechanisms to boost performance and lower the power consumption in manycore applications. This hardware-software approach is drastically different from the conventional hardware-level optimization techniques that can no longer automatically guarantee performance scalability, given the increasing diversity and complexity of manycore applications. The outcomes of this project can be used both in software development and education to more precisely understand the manycore application behavior. On a broader level, this introspective framework can boost programmer productivity by offering fast and more accurate feedback about the program execution, and effectively improving the software development cycle.

晶体管技术的快速发展与计算机体系结构的创新相结合，导致了众核处理器，它可以提供更高的软件性能。即使对于经验丰富的软件开发人员来说，获得可扩展的性能（例如，在八核处理器上将程序执行时间加速八倍）也是一项具有挑战性的任务。这是因为并行程序可能会遇到在顺序应用程序中不存在的性能问题，或者会因并行执行而恶化。此外，众核应用程序的功耗是限制其广泛采用的关键因素。不幸的是，由于程序是用多种语言和源文件编写的，库是仅二进制形式的，因此仅使用软件工具来分析众核应用的性能和功率带来了挑战。一种有效的硬件-软件框架，可以提供对软件执行的性能和功率分布的洞察，并随后促进补救机制以克服观察到的程序瓶颈，将是实现众核计算的重要一步。该研究项目探索了一种变革性的硬件-软件协作解决方案，其中软件接收对硬件行为的内省，并动态部署补救机制，以提高性能并降低众核应用程序的功耗。这种硬件-软件方法与传统的硬件级优化技术有很大的不同，传统的硬件级优化技术不能再自动保证性能可扩展性，因为众核应用程序的多样性和复杂性不断增加。该项目的成果可以用于软件开发和教育，以更准确地理解众核应用程序的行为。在更广泛的层面上，这种内省框架可以通过提供关于程序执行的快速和更准确的反馈来提高程序员的生产力，并有效地改善软件开发周期。