边缘环境下DNN应用的计算迁移调度技术

Computation Offloading Scheduling Technology for DNN Applications in Edge Environment

深度神经网络(Deep Neural Network,DNN)应用对运行设备的性能要求较高,无法直接在计算资源受限的移动设备上运行。通过计算迁移技术将某些计算复杂的神经网络层迁移到资源丰富的边缘或者远程云端上去执行,是一种有效的解决资源受限问题的方法。计算迁移会产生额外的时间开销,如果迁移过程的时延太长,将严重影响用户体验。为此,文中以得到边缘环境下多任务并行调度的最小平均响应时间为目标,首先提出边缘环境下DNN应用的计算迁移调度问题,并对该问题的解设计了评估算法;然后设计了两种调度算法即贪心算法和遗传算法(Genetic Algorithm,GA)来求解问题;最后设置评估实验,在5种不同的边缘环境下对两种算法的性能进行对比分析。实验数据表明,采用所提算法得到的解十分接近最优解。与传统的迁移方案相比,贪心算法能得到平均响应时间更短的调度方案;遗传算法的平均响应时间比贪心算法短,但其运行时间明显更长。实验结果说明,所提两种调度算法能够有效地缩短边缘环境下DNN应用的计算迁移调度的平均响应时间,提高用户体验。

Deep neural network(DNN)applications require high performance of running equipment,and can not run directly on mobile devices with limited computing resources.It is an effective method to offload some computationally complex neural network layers to resource-rich edges or remote clouds for execution by computation offloading technology.Computation offloading will incur additional time overhead.If the offloading process lasts too long,the user experience will be seriously affected.To this end,in order to obtain the minimum average response time of multi-task parallel scheduling in edge environment,this paper first proposes the computation offloading scheduling problem for DNN applications in edge environment,and designs an evaluation algorithm for the solution to the problem.Then two scheduling algorithms,greedy algorithm and genetic algorithm,are designed to solve the problem.Finally,an evaluation experiment is set up to compare and analyze the performance of the two algorithms in five different edge environments.The experimental data shows that the solution obtained by the proposed algorithms in this paper is very close to the optimal solution.Compared with traditional offloading schemes,greedy algorithm can obtain a scheduling scheme with shorter average response time.The average response time of genetic algorithm is shorter than that of greedy algorithm,but its running time is significantly longer.The experimental results show that the proposed two scheduling algorithms can effectively reduce the average response time of computation offloading scheduling for DNN applications in edge environment and improve user experience.