Joint Task Dispatching and Bandwidth Allocation with Hard Deadlines in Distributed Serverless Edge Computing Systems

Serverless computing lifts the burden of infrastructure maintenance from application developers and also reduces the usage cost of cloud/edge computing platforms. However, the pay-as-you-go pricing model offered by serverless computing complicates the task dispatching problem in serverless computing systems. Specifically, when the pay-as-you-go pricing model is adopted, the task response latency is not simply the task execution latency, but the cold-start latency and the container image downloading latency should also be considered as part of the task response latency. In this paper, we focus on the joint task dispatching and bandwidth allocation problem with hard deadlines in distributed serverless edge computing systems. To maximize the overall profit, a new algorithm called PN-GRD is presented. PN-GRD first uses a Pointer Network model that is well trained offline to inference a task permutation, which is used to determine the task priority. Then, multiple edge node selection steps are carried out to select an edge node for each task according to the task priority. The final task dispatching and bandwidth allocation is obtained once the best edge node for every task is chosen or none of the edge nodes is suitable to run a task. We validate the performance of PN-GRD through simulations. The results show that PN-GRD outperforms practical baselines in terms of the average overall profit.