CAREER: GPU Performance Portability for Volunteer Computing through Heterogeneity-aware Autotuning

该奖项部分由2021年美国救援计划法案（公法117-2）资助。计算能力对于发现引力波等现象以及加速疫苗等新药的发现越来越重要。图形处理单元（gpu）是一种经常用于加速此类科学计算的计算设备。为了达到最大的性能，软件必须调整到正在使用的特定GPU。尽管这种自动性能调优（autotuning）在所有gpu都相同的传统超级计算环境中是常规和有效的，但在包含数百种不同类型gpu的环境中，该过程不能很好地扩展或提取最大性能。这种异构性在志愿者计算中很常见，在志愿者计算中，捐赠的计算机时间用于执行大量计算。通过开发能够处理异构的自动调整算法，向程序员提供有关未实现性能的反馈，并在真正的志愿计算系统中大规模工作，该项目将使这些系统能够最大化GPU性能，并加速医学、生物学和天文学等领域的科学发现。该项目的目标是自动调整应用程序，以便在高度异构的志愿计算环境中以较低的成本在所有GPU上获得最大性能，而不是在遇到的每个GPU上重新运行自动调优。首先，开发了识别gpu之间相似性的方法，以减少自动调优的工作量，同时使用跨不同gpu收集的性能数据来引导和加速自动调优过程。其次，性能模型与自动调优集成，向程序员提供有关瓶颈和错过的性能优化机会的反馈。第三，并行和分布式搜索用于减少应用程序性能方面的非生产性探索。所有这些都将在现实世界流行的BOINC志愿计算系统中实现，以使现有和未来的科学志愿计算项目受益。该项目将为K-12学生开展外展活动，并通过提供高保真GPU模型来帮助保存电子游戏。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This award is funded in part under the American Rescue Plan Act of 2021 (Public Law 117-2).Computational power is increasingly critical for discovering phenomena such as gravitational waves and to accelerate the discovery of new drugs such as vaccines. Graphics Processing Units (GPUs) are computational devices that are frequently used to speed up such scientific calculations. To achieve maximum performance, software must be tuned to the particular GPU being used. Although such automatic performance tuning (autotuning) is routine and effective in conventional supercomputing environments where all the GPUs are identical, the process does not scale well or extract maximum performance in environments containing hundreds of different types of GPUs. Such heterogeneity is common in volunteer computing where donated computer time is used to perform massive computations. By developing autotuning algorithms that can deal with heterogeneity, providing feedback to programmers about unrealized performance, and working at scale in real volunteer computing systems, this project will enable these systems to maximize GPU performance and accelerate scientific discovery in fields such as medicine, biology, and astronomy. The goal of this project is to automatically adapt applications to obtain maximal performance on all GPUs in the highly heterogeneous volunteer computing environment at lower cost than re-running autotuning on each GPU encountered. First, methods to identify similarities among GPUs are developed to reduce autotuning effort while performance data gathered across different GPUs is used to bootstrap and speed up the autotuning process. Second, performance models are integrated with autotuning to provide feedback to programmers about bottlenecks and missed performance optimization opportunities. Third, parallel and distributed search is used to prune unproductive explorations of an application’s performance landscape. All these will be implemented in the popular real-world BOINC volunteer computing system to benefit both existing and future scientific volunteer computing projects. The project will conduct outreach activities for K-12 students, and also aid the preservation of electronic games by providing hi-fidelity GPU models.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.