【作者】 李娜；

【导师】 高博；

【作者基本信息】 郑州大学 ， 工程硕士（专业学位）， 2022， 硕士

【摘要】 随着通信技术和异构计算的飞速发展,为满足信号处理领域大计算、高吞吐、超大带宽、超低时延等需求,异构信号处理平台成为该领域的研究热点。而平台系统性能与应用任务部署紧密相关,因此面向此类平台的任务部署算法已成为解决信号处理领域的关键问题之一。它根据任务多样性需求分配合适的处理、通信资源,以实现不同任务的并行处理,提高任务的执行效率,满足实时性要求。本文依托某“十四五”研究项目,围绕异构信号处理平台的任务部署算法展开研究,主要研究内容如下:(1)论文通过研究异构信号处理平台软硬件架构,分析信号处理任务组件化思想,对组件和平台硬件进行DAG抽象,建立数学模型;结合平台资源约束和任务实际需求,定义任务类型及异构平台下任务部署相关参数,提出相关性能评价指标,为后续算法对比提供理论支撑。(2)针对计算密集型任务采用传统蚁群算法部署存在搜索空间大、收敛速度慢的问题,提出了基于强化学习的改进蚁群任务部署算法。计算密集型任务具有良好的并行特性,达到最佳调度长度是实现任务高效执行的关键。论文首先将抽象后的组件和处理器模型与Q-learning算法进行场景适配,描述任务部署状态空间、动作空间,建立奖励矩阵,通过多次迭代得到Q矩阵;然后将Q矩阵作为蚁群算法的初始信息素提高搜索效率,结合任务排序列表和状态转移概率实现任务到处理器的映射。实验表明,与传统蚁群算法(ACO)、改进遗传算法(GA)、QMTS算法相比,本文所提ACOQ算法在计算密集型任务中调度长度平均分别减少38.30%、22.69%、21.35%。(3)针对通信密集型任务通信开销较大,部署过程易导致系统负载不均衡、限制平台性能的问题,提出了基于边聚簇的任务复制部署算法。论文首先通过任务划分和边聚簇压缩通信规模,减少通信量和计算复杂度;在此基础上,通过多任务复制技术充分利用计算单元的空闲时间片,并进一步减小组件间的通信时延,同时通过动态负载均衡达到平台资源的合理利用。实验表明,与HEFT、ACOQ、LS-IPLB算法相比,本文所提HCLA算法在负载均衡方面得到提高;HEFT和ACOQ算法在CCR较大时加速比小于1,不适用于通信密集型任务,而HCLA和LS-IPLB算法加速比均大于1,且HCLA相比LS-IPLB算法加速比提高8.9%,更适合通信密集型任务。(4)为实现异构信号处理平台任务部署模型及算法的直观展示和对比,论文利用PYQT的逻辑与界面分离特性,基于信号和槽机制设计了多个模块,集成异构信号处理平台任务部署可视化演示系统,为平台任务调度提供支撑。更多还原

【Abstract】 With the rapid development of communication technology and heterogeneous computing,heterogeneous signal processing platform has become a research hotspot in this field in order to meet the needs of large computing,high throughput,super bandwidth and ultra-low delay in the field of signal processing.The performance of platform system is closely related to application task deployment,so the task deployment algorithm for this kind of platform has become one of the key problems in the field of signal processing.It allocates appropriate processing and communication resources according to the requirements of task diversity,so as to realize the parallel processing of different tasks,improve the execution efficiency of tasks and meet the requirements of real-time.Relying on a "14th Five-Year Plan" research project,this paper focuses on the task deployment algorithm of heterogeneous signal processing platform.The main research contents are as follows:(1)Aiming at the heterogeneous signal processing platform,this paper studies its software and hardware architecture,analyzes the component idea of signal processing task,DAG abstracts the components and platform hardware,and establishes the mathematical model.Combined with the platform resource constraints and the actual needs of tasks,the types of tasks and the relevant parameters of task deployment under heterogeneous platforms are defined,and the relevant performance evaluation indexes are proposed to provide theoretical support for the comparison of subsequent algorithms.(2)For computing intensive tasks,ant colony algorithm has the problems of large search space and slow convergence speed to schedule tasks.An improved ant colony task deployment algorithm based on reinforcement learning is proposed.Computing intensive tasks have good parallel characteristics,and achieving the optimal scheduling length is the key to achieve efficient task execution.The paper first adapts the abstract task and processor model to the scene of Q-learning algorithm,describes the task deployment state space and action space,establishes the reward matrix,and obtains the Q matrix through multiple iterations;Then the Q matrix is used as the initial pheromone of ant colony to improve the search efficiency,mapping tasks to processors combined with task scheduling list and state transition probability.Experiments show that the proposed ACOQ algorithm in computing intensive tasks reduces the scheduling length by 38.30%,22.69% and 21.35% respectively compared with the traditional ant colony algorithm(ACO),improved genetic algorithm(GA)and QMTS algorithm.(3)For communication intensive tasks,its communication sales are large and the deployment process is easy to lead to system load imbalance and limit the performance of the platform,a task replication deployment method based on edge clustering is proposed.The communication overhead of communication intensive application tasks is large,data transmission delay and system load has become the key to limit the performance of the platform.Firstly,the communication scale is compressed by task division and edge clustering to reduce the communication overhead and computational complexity.On this basis,multi task replication technology makes full use of the idle time slice of the computing unit,reduces the communication delay between components,and achieves the rational utilization of platform resources through dynamic load balancing.Experiments show that compared with HEFT,ACOQ and LSIPLB,this algorithm(HCLA)performs better in load balancing;The speedup ratio of HEFT and ACOQ algorithms is less than 1 when the CCR is large,so they are not suitable for communication intensive tasks;The speedup ratio of HCLA and LS-IPLB algorithm is greater than 1,and the speedup ratio of HCLA is 8.9% higher than that of LS-IPLB,which is more suitable for communication intensive tasks.(4)In order to realize the visual display of task deployment model and algorithm of heterogeneous signal processing platform,this paper uses the logic and interface separation characteristics of PYQT,designs multiple modules based on the signal and slot mechanism to integrate the task deployment visual demonstration system of heterogeneous signal processing platform,so as to provide support for platform task scheduling.更多还原