Air Option 1: Technology Translation - Compiler Technology for Modern Manycore Architectures

Air 选项 1：技术翻译 - 现代众核架构的编译器技术

• 批准号: 1343436
• 负责人: Aviral Shrivastava
• 金额: $ 15万
• 依托单位: Arizona State University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2013
• 资助国家: 美国
• 起止时间: 2013-09-15 至 2016-02-29
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1343436&HistoricalAwards=false
• 关键词: Air; Option; Technology; Translation; Compiler

This AIR Technology Translation project focuses on translating manycore compiler technology to fill the designer productivity and manycore usage efficiency gap.There is a wide consensus among computer architects that existing multicore processor designs do not scale due to the complexity of memory hierarchy. As a result memory hierarchies in manycore architectures are expected to be simpler ? with the onus of memory management now in software. However, doing that is extremely complicated for the programmer. The proposed manycore compiler technology alleviates this programmer difficulty. It automatically inserts data management instructions in the application and enables their efficient execution on manycore processors.The translated technology has the following unique features: push-button compilation and automatic yet efficient data management that provides exemplary improvements in software development time and cost for near future manycore processors. In addition, it will also result in improved performance, and lower power consumption of the application executing on the manycore processor.The project accomplishes this goal by developing a LLVM based compiler that will analyze the application and insert data management instructions in the program, so that the applications execute correctly and efficiently, resulting in a prototype compiler that will enable the execution of existing applications on manycore processors. While there are compiler techniques that solve a part of the problem, but our compiler not only outperforms them [see the results in our recent publications], but provides a holistic solution to the manycore programming challenge.The partnership engages Intel (a manycore processor development company) and Raytheon Missile Systems (a manycore software development company) to provide guidance in the manycore processor and software marketplace and other aspects, including financing and commercialization as they pertain to the potential to translate the technology along a path that may result in a competitive commercial reality.The potential economic impact of our manycore compiler technology is expected to be very significant, as this compiler technology will be needed by companies that want to use manycore processing for high performance and low power. For example, 3D tomography, high capacity 3D printers, ultra high definition 3D TVs, automatic target recognition and collision avoidance at supersonic speeds, etc., all need highest levels of performance, under strict power constraints. The power and performance needs of these applications can only be met by manycore processors, but manycore processors are not expected to have the support for memory management in hardware. Therefore existing applications will not execute on them. Our compiler will enable the execution of these applications, and therefore utilize the power and performance of latest manycore processors. This edge can greatly contribute to the U.S. competitiveness in information technology, defense systems, and consumer electronic market space.The societal impact, in the long term, will be that the end user will receive high performance, yet power-efficient computing products at reasonable costs. This improves affordability of computing products that we need for everyday living, defense, services, healthcare etc., and elevates standard of living, especially for the economically disadvantaged.

AIR技术翻译项目致力于翻译多核编译器技术，以填补设计人员生产力和多核使用效率的差距。计算机架构师普遍认为，由于内存层次结构的复杂性，现有的多核处理器设计无法扩展。因此，预计多核体系结构中的内存层次结构会更简单？内存管理的责任现在由软件来承担。然而，对于程序员来说，这样做是非常复杂的。提出的多核编译器技术缓解了程序员的这一困难。它自动将数据管理指令插入到应用程序中，并使其能够在许多核心处理器上高效执行。翻译后的技术具有以下独特特征：按下按钮编译和自动但高效的数据管理，为不久的将来的多核处理器提供软件开发时间和成本方面的示范性改进。该项目通过开发一个基于LLVM的编译器来实现这一目标，该编译器将分析应用程序并在程序中插入数据管理指令，从而使应用程序能够正确高效地执行，从而产生能够在多核处理器上执行现有应用程序的原型编译器。虽然有一些编译器技术可以解决部分问题，但我们的编译器不仅性能优于它们(参见我们最近出版物中的结果)，而且为多核编程挑战提供了整体解决方案。该合作伙伴关系与英特尔(多核处理器开发公司)和雷神导弹系统公司(多核软件开发公司)合作，在多核处理器和软件市场和其他方面提供指导，包括融资和商业化，因为它们与将技术转化为可能导致竞争性商业现实的道路的潜力有关。我们的多核编译器技术的潜在经济影响预计将非常显著，因为这种编译器技术将被希望使用多核处理以实现高性能和低功耗的公司所需要。例如，3D断层扫描、高容量3D打印机、超高清3D电视、自动目标识别和超音速碰撞避免等，都需要最高级别的性能，在严格的功率限制下。这些应用程序的能力和性能需求只能由多核处理器来满足，但多核处理器预计不会在硬件中支持内存管理。因此，现有的应用程序将不会在它们上执行。我们的编译器将支持这些应用程序的执行，从而利用最新的多核处理器的能力和性能。这一优势可以极大地增强美国在信息技术、国防系统和消费电子市场空间的竞争力。从长远来看，社会影响将是最终用户将以合理的价格获得高性能、高能效的计算产品。这提高了我们日常生活、国防、服务、医疗等所需的计算产品的可负担性，并提高了生活水平，特别是对经济困难的人来说。