Deep learning applications executed on mobile devices can provide artificial intelligence services with multi-access edge computing (MEC). Most of existing works on computation offloading neglect the context switch cost, which is one of the dominating factors that affect the finishing time of deep learning applications. This paper thus tries to investigate the context switch cost aware joint task merging and scheduling for deep learning applications. We formulate the problem as a non-linear integer programming, and prove that it is NP-Hard. The objective is to minimize the finishing time of deep learning applications with energy budget constraints. To solve the problem, an efficient joint merging and scheduling algorithm named fusion merging scheduling (FMS) is proposed to fully exploit the potential influence of context switch cost in distributed execution of deep learning tasks. A real-world platform is built to evaluate the performance of proposed algorithm. Experimental results show that FMS can reduce the average finishing time of deep learning applications by 79%, 63% and 75%, respectively, compared with a greedy scheduling algorithm and two existing benchmarks, while achieving similar energy costs.

在移动设备上执行的深度学习应用程序可以为人工智能服务提供多访问边缘计算（MEC）。现有的大部分计算卸载工作都忽略了上下文切换成本，这是影响深度学习应用程序完成时间的主要因素之一。因此，本文试图研究深度学习应用的上下文切换成本意识联合任务合并和调度。我们将该问题表述为非线性整数规划，并证明它是NP-Hard。目的是在能源预算约束下最大程度地缩短深度学习应用程序的完成时间。为了解决这个问题 为了充分利用上下文切换成本在深度学习任务的分布式执行中的潜在影响，提出了一种有效的联合合并与调度算法，称为融合合并调度（FMS）。建立了一个现实世界的平台来评估所提出算法的性能。实验结果表明，与贪婪调度算法和两个现有基准相比，FMS可以将深度学习应用程序的平均完成时间分别减少79％，63％和75％，同时实现相似的能源成本。