Joint task offloading and computing resource allocation with DQN for task-dependency in Multi-access Edge Computing

The rapid development of mobile communication networks has given birth to various computation-intensive applications such as augmented/virtual reality. Multi-access Edge Computing (MEC), which offloads the computation tasks of Internet of Things (IoT) devices to edge servers near terminals, has been regarded as an effective approach to achieve efficient computing offloading and reduce the heavy computation burdens. However, edge servers typically only have limited resources, which are competed and shared by IoT devices. Most of the existing researches on resource allocation focus on independent tasks, which is difficult to meet the challenge in real applications consisting of multiple interdependent tasks. In this paper, we study joint task offloading and computing resource allocation in task-dependent MEC systems. To evaluate this problem, we formulate this problem to minimize the weighted sum of the long-term task execution delay and energy consumption of IoT devices, which takes the maximum tolerable delay of the task as an important constraint. Since the problem is NP-hard, we design a joint task offloading and computing resource allocation algorithm based on Deep Q-Network (DQN) for multi-user and multi-dependent tasks (JTOCRA-DQN). Different from the traditional DQN algorithm, we add a multi-user multi-dependent task joint task offloading and computing resource allocation algorithm preparation step before learning to reduce the action space. Extensive simulation experiments show that JTOCRA-DQN can reduce the total cost by nearly 30% compared with other methods. When the maximum tolerance time is 10 s, the total cost of JTOCRA-DQN is 72.8, which is 21.3% lower than that of RoFFR algorithm and 36.8% lower than that of DREAM algorithm. In terms of the balance between energy consumption and delay, when the energy threshold increases from 70% to 95%, the proportion of energy cost in the total cost reduces from 58% to 43%, and the proportion of delay cost increases from 42% to 57%, which reflects the dynamic trade-off between energy consumption and delay.