SHF: Medium: Energy Efficient Computing on GPU-based Heterogeneous Systems

当前设计未来多处理器的趋势是在单个平台上集成数百个内核和硬件加速器(HAD)，例如GPU。然而，随着系统规模的扩大，这些异构型多处理器的功耗和能耗远远超出了预算。该项目的主要重点是开发可在与CPU适当协调的情况下在GPU中实施的节能技术。通过将当前的CPU技术扩展到GPU，可以在基于GPU的异构型多处理器中实现节能。该项目为GPU核心缩放和动态电压和频率缩放(DVFS)开发了新的运行时技术，并将它们与算法和数据结构的基本更改相结合，以提高能源效率。现有的性能和能耗模型假定处理核心的利用率为100%，并与执行期间的内存访问完美重叠。本项目针对不同的应用，综合考虑了算法、数据结构、缓存和内存聚合等因素，开发了更加准确的能效模型。正在开发一个运行时系统，该系统监控GPU内核和内存利用率以及执行应用程序时的能源消耗。运行时通过预测动态调整核心数量和/或频率级别，但随着执行的进行继续进行修正。运行时间被扩展到由CPU和GPU组成的异类系统。目前用于科学计算应用的DVFS技术不能完全消除冗余，因此不是能量最优的。利用算法的特点，开发了一种线性代数应用程序的频率调度技术，以达到更好的能效，该项目在划分和设计应用程序的任务时优化了能效。在保持通用性的前提下，以Cholesky因式分解为例，设计了一种节能调度器。该项目开发的软件产品可以很容易地应用于现有的大型异类计算机，执行适合国防、能源和关键基础设施项目的科学应用程序。此外，还开发了天气预报和结构动力学等对社会产生重大影响的应用程序。研究内容被整合到研究生课程中，为学生提供设计和编程不同类型系统的培训。该项目旨在培养包括女学生在内的高素质博士毕业生。加州大学河滨分校以其很大比例的拉美裔学生而闻名，而加州大学河滨分校是一所面向少数族裔的机构。该项目支持招收代表性不足的少数族裔和女性学生。

The current trend in designing future multiprocessors is to integrate hundreds of cores and hardware accelerators (HAs), such as GPUs, on a single platform. However, as the system size scales, the power and energy consumption of these heterogeneous multiprocessors vastly exceed the budget. The primary focus of the project is to develop energy efficient techniques that can be implemented in the GPU with proper coordination with the CPUs. Energy reduction in GPU-based heterogeneous multiprocessors can be achieved by extending current CPU techniques to GPU. The project develops new runtime techniques for GPU core scaling and Dynamic Voltage and Frequency scaling (DVFS), and combines them with basic changes in algorithms and data structures to improve energy efficiency. Existing models for performance and energy consumption assume 100% utilization of the processing cores and perfect overlap with memory access during execution. This project develops a more accurate model for energy efficiency taking into account the algorithm, data structure, caching and memory coalescing for different applications. A runtime system is being developed that monitors the GPU core and memory utilizations together with the energy consumption while executing an application. The runtime adjusts the number of cores and/or frequency level dynamically through prediction, but continues to make corrections as the execution proceeds. The runtime is extended to heterogeneous systems consisting of both CPU and GPU.The current DVFS techniques for scientific computing applications cannot fully eliminate slacks, therefore, are not energy optimal. By leveraging the algorithmic characteristics, a frequency scheduling technique is developed for linear algebra applications to achieve better energy efficiency.The project optimizes the energy efficiency while partitioning and designing tasks of an application. Without loss of generality, Cholesky factorization is used as an example and an energy efficient scheduler is developed. The project develops software products that can be readily applied to existing large scale heterogeneous computers executing scientific applications that are suitable for defense, energy and critical infrastructure projects. Also applications like weather forecasting and structural dynamics are developed that have a great impact on society. The research content is integrated to graduate courses to provide training to students for designing and programming heterogeneous systems. The project aims to produce very high quality Ph.D. graduates including female students. The University of California, Riverside is known for its large proportion of Hispanic students, and UCR is a minority-serving institution. The project supports recruiting underrepresented minority and female students.