Dynamic Task Allocation for Cost-Efficient Edge Cloud Computing

Edge cloud computing systems are widely used to supply various computation services in Internet of Things (IoT). An essential problem is how to efficiently allocate task requests to various edge and cloud servers given task requirements (e.g., response time and required memory space), in order to minimize various costs generated in edge cloud computing. Existing studies on task allocation usually consider the viewpoint of provider cost such as offloading cost, uploading cost and deployment cost. However, the viewpoint of user cost (e.g., server fee) is rarely considered which is becoming an important issue in the deployment of edge cloud computing systems, especially for cost sensitive users like venture companies. In this paper, we study a dynamic task allocation problem in edge cloud computing where both servers' status and arriving tasks would change along with time; the goal is to search the task allocation policy that can minimize user cost. Specifically, we consider a parallel processing case where a task's workload can be infinitely divided among the various servers; this causes a huge solution space and makes the problem hard to solve. Thus, we consider an approximate method from the perspective of server coalitions rather than a single server, and propose a dynamic coalition formation algorithm called coalitional R-learning (CR-learning) to guide several edge servers in forming a coalition dynamically. Simulations verify that our algorithm can significantly reduce user cost comparing with some other existing algorithms while shrinking the solution space.