Performance for All: Broadening the Scope of Parallelism and Specialization

所有人的性能：扩大并行性和专业化的范围

• 批准号: RGPIN-2022-05330
• 负责人: Jeffrey, Mark
• 金额: $ 2.11万
• 依托单位: University of Toronto
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2022
• 资助国家: 加拿大
• 起止时间: 2022-01-01 至 2023-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=751601
• 关键词: Performance; Broadening; Scope; Parallelism; Specialization

Energy efficiency is the critical challenge facing modern computer architectures. Demand for higher performance will continue its historic growth, yet the technology scaling trends that backed decades of performance and efficiency improvements are now flat-lining. To meet user demand and continue performance scaling, architectures require new ways to reduce energy per operation. Parallelism (running many operations simultaneously) and specialization (customizing circuitry and data movement for a given application domain or set of patterns) are two well-known and complementary techniques to improve efficiency. A small minority of application domains have inspired significant investment into specialized architectures (e.g., Google's TPU), largely enabled by the enormous economic value of their domains, and "embarrassing" amounts of data parallelism which simplify hardware design. Unfortunately, for the vast majority of applications, the fixed cost of fabricating custom hardware is prohibitive, and in general they have more complex forms of irregular parallelism that are a poor match to modern multicores, GPUs, and other accelerators. These applications currently have no pathway to future performance gains, limiting the size and scope of problems that can be solved. This research program takes a holistic approach to improve performance for all by co-designing programming models, multicores, reconfigurable accelerators, compilers, and languges for the heavy tail of these underserved applications. We will solve compute bottlenecks in application domains such as statistical inference, graph analytics, electronic CAD, mathematical optimization, and more. Through the program, the PI will train highly-qualified personnel (HQPs) in the areas of computer architecture and systems which are sought-after in the technology industries within and outside Canada. Particular skills include performance analysis, compiler optimization, system simulation and prototyping, and hardware-software co-design. The PI will address equity, diversity, and inclusion (EDI) in HQP recruitment and the training environment. The PI will advertise broadly and recruit proactively to attract a diverse applicant pool. The PI will routinely emphasize EDI principles during training, and will raise HQP profiles by connecting those from under-represented minorities to groups such as Women in Computer Architecture or Black in AI.

能源效率是现代计算机体系结构面临的关键挑战。对更高性能的需求将继续保持其历史性的增长，但几十年来支持性能和效率提高的技术扩展趋势现在趋于平缓。为了满足用户需求和持续的性能扩展，架构需要新的方法来减少每次操作的能耗。并行（同时运行许多操作）和专门化（为给定的应用程序领域或模式集定制电路和数据移动）是提高效率的两种众所周知的互补技术。一小部分应用领域激发了对专门架构（例如b谷歌的TPU）的大量投资，这主要是由于其领域的巨大经济价值，以及简化硬件设计的“令人尴尬”的数据并行性。不幸的是，对于绝大多数应用程序来说，制造定制硬件的固定成本是令人望而却步的，而且通常它们具有更复杂的不规则并行形式，与现代多核、gpu和其他加速器不太匹配。这些应用程序目前没有办法提高未来的性能，限制了可以解决的问题的规模和范围。该研究项目采用整体方法，通过共同设计编程模型、多核、可重构加速器、编译器和针对这些服务不足的应用程序的沉重尾部的语言，来提高所有人的性能。我们将解决诸如统计推断、图形分析、电子CAD、数学优化等应用领域的计算瓶颈。通过该计划，PI将在加拿大国内外的技术行业中培养计算机体系结构和系统领域的高素质人才（HQPs）。特殊技能包括性能分析，编译器优化，系统模拟和原型设计，以及硬件软件协同设计。PI将解决HQP招聘和培训环境中的公平、多样性和包容性（EDI）问题。PI将广泛宣传并积极招聘，以吸引多样化的申请人。PI将在培训期间定期强调EDI原则，并将通过将未被充分代表的少数民族与计算机架构中的女性或人工智能中的黑人等团体联系起来，提高HQP的形象。