CPU/GPU系统负载均衡的可分负载调度  

Load-balanced scheduling for divisible workloads on hybrid CPU/GPU systems

作　　者：彭江泉[1] 钟诚[1] 

机构地区：[1]广西大学计算机与电子信息学院,广西南宁530004

出　　处：《计算机工程与设计》2013年第11期3916-3923,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(60963001)

摘　　要：针对具有多个片上多核处理器CMP和多个GPU的异构系统,综合考虑多核CPU和GPU的不同计算能力、各级缓存不同容量、CPU和GPU之间通信代价以及GPU内部结构等因素,采取计算与通信重叠、对任务自动划分、GPU端线程块大小和维度自动设置、多个异步流同时传输的方法,通过切换线程块执行以隐藏访存开销,提出一种均衡CPU和GPU负载、高效的可分负载多轮调度算法。采用计算密集型任务Mandelbrot进行实验测试,测试结果表明,与已有算法相比,该算法的调度性能有明显提升。Considering the factors such as the different computation abilities for CPU and GPU, distinct capacity for L3 cache, I2 cache and L1 cache, communication cost between CPU and GPU, and the internal structure of GPU, an efficient load-balanced multi-round scheduling algorithm for divisible loads is proposed on hybrid multi-core CPU and GPU systems with multiple CMPs and multiple GPUs by applying the overlapping of computation and communication, automatic partitioning tasks, setting the size and dimension of threacg block, simultaneously transmitting multiple asynchronous streams, and switching the execution of thread-block to hide memory access overhead. The experimental results with the compute-intensive task Mandelbrot show that compared with the existing algorithm, the presented scheduling algorithms can obtain a large performance enhancement.

关 键 词：CPU GPU异构系统 可分负载 调度算法 协同计算 负载均衡 多级缓存 

分 类 号：TP338.6[自动化与计算机技术—计算机系统结构] TP301.6[自动化与计算机技术—计算机科学与技术]