高级检索

    CPU-GPU MPSoC中使用寿命驱动的OpenCL应用调度方法

    Lifetime-Driven OpenCL Application Scheduling on CPU-GPU MPSoC

    • 摘要: 近年来,集成CPU和GPU的多处理器片上系统(multiprocessor system-on-chips,MPSoC),凭借兼顾GPU核心的并行计算能力和CPU核心的通用计算能力,已经广泛应用于工业控制、汽车电子、智慧医疗等领域. 为了充分发挥CPU-GPU MPSoC的性能,开放计算语言(open computing language,OpenCL)逐渐成为一种主流的应用程序编写标准.然而,在将OpenCL应用部署到CPU-GPU MPSoC的过程中,现有研究工作大多忽略了对芯片温度和使用寿命的管理,导致处理器核心在执行应用时超过了峰值温度,甚至永久性故障的提前发生,无法保证OpenCL应用的长久稳定运行. 为了弥补上述缺点,提出了一种包含静态和动态应用调度技术的方法. 静态应用调度技术是基于改进交叉熵策略,将OpenCL应用的特性充分考虑在内,有效提高了OpenCL应用设计点的寻优效率. 动态应用调度技术是基于反馈控制策略,克服了传统方案中无法有效应对系统运行时新到应用的缺陷,能够最小化新到应用的平均延迟. 实验表明, 所提方法可以将应用的平均延迟降低34.58%,同时满足温度、能耗、使用寿命的约束.

       

      Abstract: In recent years, multiprocessor system-on-chips (MPSoC) integrating CPU and GPU have been widely deployed in the fields of industrial control, automotive electronics, smart medical, etc. The open computing language (OpenCL) is regarded as a popular application programming standard for CPU-GPU MPSoC due to the power of fully exploiting the parallel computing power of GPU cores and the general-purpose computing power of CPU cores. However, during deploying OpenCL applications to CPU-GPU MPSoC, most of the existing research works have neglected the management of chip temperature and lifetime, resulting in the elevated peak temperature and the early occurrence of permanent failures. In this paper, we explore the lifetime-driven OpenCL application scheduling for latency minimization on CPU-GPU MPSoC under timing, temperature, energy consumption, and lifetime constraints. We propose a method composed of static and dynamic application scheduling techniques. The static application scheduling technique is built on the improved cross-entropy strategy with consideration of the OpenCL application characteristics in searching for optimal OpenCL application design points. The dynamic application scheduling technique is developed on the feedback control strategy capable of processing the new arrival applications for latency optimization at runtime. Experimental results show that our proposed method reduces the average delay of OpenCL applications by 34.58% while satisfying all design constraints.

       

    /

    返回文章
    返回