CPU-GPU异构高性能计算中的负载预测调度算法研究及应用

发布时间：2018-04-21 03:03

本文选题：负载预测调度算法 + CPU　；参考：《上海大学》2016年博士论文

【摘要】：由于性价比和能效比很高,多核CPU-GPU计算平台得到了广泛应用,这也使系统内同时存在两种异构的计算资源。但是多核CPU和GPU的性能必须通过高效的调度算法才能得到充分发挥。因此如何充分利用异构资源的计算能力,如何实现负载均衡成为研究的热点。传统的调度方法有静态调度和动态调度。静态调度开销非常小,但容易导致负载不均衡,降低计算资源的利用率;动态调度能更好地实现负载均衡,但调度开销比较大。如果将上述两种调度方法结合起来,将大大减少调度开销,并有效地实现负载均衡。CPU-GPU异构计算平台中,基于SIMD结构的GPU适合并行度和计算量大的计算任务,GPU的计算性能远远大于CPU的计算性能,但是现有的调度算法无法根据硬件特点进行任务分配。本文针对上述问题,提出了一种新的调度方法--负载预测调度算法(Load-prediction scheduling--LPS),该算法可以充分发挥异构的多核CPU和GPU的计算能力,并实现静态和动态调度的有效结合。本文完成的主要工作包括:1、本文提出了负载预测调度算法,该算法具有以下特点:(1)根据GPU硬件特点分配任务,充分发挥GPU计算性能。(2)有效结合了动态调度和静态调度,实现了负载均衡和减少调度开销,适合应用在异构环境中。(3)充分发挥多核CPU的计算性能。2、将负载预测调度算法应用在心电仿真计算中,实现了上述特点。此外在计算中通过负载预测消除了分支,同时提高了GPU的计算粒度,因此拓宽了GPU的计算范围,进一步提高了计算效率。3、将负载预测调度算法应用在多体问题计算中。
[Abstract]:Multi-core CPU-GPU computing platform has been widely used because of its high performance-to-price ratio and energy-efficiency ratio, which also makes two heterogeneous computing resources exist in the system at the same time. But the performance of multi-core CPU and GPU must be achieved by efficient scheduling algorithm. Therefore, how to make full use of the computing power of heterogeneous resources and how to achieve load balancing has become a hot topic. Traditional scheduling methods include static scheduling and dynamic scheduling. Static scheduling overhead is very small, but it is easy to lead to load imbalance, reduce the utilization of computing resources, dynamic scheduling can better achieve load balancing, but scheduling overhead is relatively large. If the above two scheduling methods are combined, the scheduling overhead will be greatly reduced, and the load balancing. CPU-GPU heterogeneous computing platform will be implemented effectively. GPU based on SIMD structure is suitable for computing tasks with high degree of parallelism and large amount of computation. The performance of GPU is much higher than that of CPU, but the existing scheduling algorithms can not assign tasks according to the hardware characteristics. In this paper, a new scheduling method, Load-Prediction scheduling algorithm, is proposed in this paper. The algorithm can give full play to the computing power of heterogeneous multi-core CPU and GPU, and realize the effective combination of static and dynamic scheduling. The main work accomplished in this paper includes: 1. This paper proposes a load prediction scheduling algorithm. The algorithm has the following characteristics: 1) assign tasks according to the hardware characteristics of GPU, give full play to the performance of GPU computing. 2) effectively combine dynamic scheduling with static scheduling. It realizes load balancing and reducing scheduling overhead. It is suitable for application in heterogeneous environment. It can give full play to the computing performance of multi-core CPU. The load predictive scheduling algorithm is applied to ECG simulation, and the above characteristics are realized. In addition, the branch is eliminated by load prediction, and the computational granularity of GPU is improved. Therefore, the computational range of GPU is widened, and the computational efficiency is further improved. The load prediction scheduling algorithm is applied to the computation of multi-body problem.
【学位授予单位】：上海大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP301.6

【参考文献】