通用图形处理器线程调度优化方法研究综述  

Thread Scheduling Optimization of General Purpose Graphics Processing Unit:A Survey

作　　者：何炎祥[1,2] 张军[1,3] 沈凡凡[1] 江南[1,4] 李清安[1,2] 刘子骏[1] 

机构地区：[1]武汉大学计算机学院,武汉430072 [2]武汉大学软件工程国家重点实验室,武汉430072 [3]东华理工大学软件学院,南昌330013 [4]湖北工业大学计算机学院,武汉430068

出　　处：《计算机学报》2016年第9期1733-1749,共17页Chinese Journal of Computers

基　　金：国家自然科学基金(61170022,61373039,61462004)资助

摘　　要：随着通用图形处理器(GPGPU)并行计算能力的日益增强,其应用范围越来越广.然而由于不规则计算任务使得通用图形处理器资源难以得到充分利用,其性能并未达到最大化.在论述GPGPU微体系结构的基础上,该文重点讨论了多种针对GPGPU性能提升的线程调度优化方法,主要从针对优化分支转移提升线程级并行度、针对访存效率的提升和针对标量指令执行以提升资源利用率三方面的线程调度优化方法进行了分析和比较.另外,由于功耗问题已成为制约GPGPU发展的主要因素之一,该文还分析了影响GPGPU功耗效率的主要原因及目前提高GPGPU功耗效率的主要低功耗技术,并对目前针对GPGPU功耗优化的典型线程调度优化方法进行了分析比较.最后,该文指出了未来线程调度优化方法需要进一步探讨的一些问题.Due to their massive parallel computational power, General Purpose Graphics Processing Units （GPGPUs） have become an increasingly popular platform for executing general purpose applications. However the existed irregular workloads make the resource of GPU underutilized and its performance not maximized. A variety of thread scheduling optimization strategies on the performance enhancement of GPGPU are mainly discussed based on the microarchitecture of GPGPU. Specifically, the thread scheduling optimization strategies on branch divergence optimi-zation to improve thread level parallelism, memory access enhancement and scalar instruction execution to increase the resource utilization are analyzed and compared. In addition, the primary causes affecting the power are analyzed and the major methods improving the efficiency of power as well as typical thread scheduling strategies on power optimization are compared as the issue of power has become one of the bottlenecks limiting the development of GPGPU. Finally, it summarizes some research interests need to be discussed and studied in thread scheduling optimization in future.

关 键 词：通用图形处理器 线程调度优化 性能 功耗 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]