A Joint Optimization Method Of Resource Allocation And Task Scheduling For Computing-aware Netwoking

With the update and iteration of computing devices and the continuous expansion of computing scale,the concept of computing-aware networking is proposed by the enterprise and academia in order to meet the large demand for computing power in the digital transformation of industry.Based on ubiquitous network connectivity,it aggregates heterogeneous computing resources such as scattered edge clouds and third-party devices in the network,and expects to achieve global optimization of multidimensional resource allocation based on unified and collaborative management and scheduling of computing and network resources,improve the overall utilization of computing resources,and provide high-quality and low-latency services to users.Efficient multidimensional resource allocation is one of the key technologies to achieve the goal of computing-aware networking.Although current research has achieved certain results,it still needs to be developed,On the one hand,most of the literature focuses on short-term goals,but user needs change dynamically over time,and long-term goals obviously fit the actual needs better than short-term goals.On the other hand,most of the literature builds task arrival models based on theories,however,there are some deviations between theoretical models and real situations,which makes it difficult for the proposed methods to achieve optimal results in real environments.To address the above issues,this paper investigates the joint optimization method of resource allocation and task scheduling from a long-term perspective based on user access data set modeling,with latency and profit as two optimization dimensions in multiedge cloud scenario and third-party device sharing scenario.(1)Delay-aware resource allocation and task scheduling joint optimization method under multi-edge cloudIn the scenario of geographically dispersed deployment of edge clouds,improper resource allocation and task scheduling strategies will cause problems such as higher task latency and larger network traffic.Therefore,the paper proposes delay-aware joint optimization methods for resource allocation and task scheduling to reduce task response latency and backbone network traffic.First,the long-term problem is modeled as a mixed-integer nonlinear programming.Second,for the challenge of large feasible domain and difficult solution due to multi-problem coupling,a two-step decision-based algorithm framework is designed to decouple the problem into two sub-problems.Then,for the first sub-problem,a deep reinforcement learning-based task scheduling method is designed to obtain a task scheduling policy.The second subproblem is proved to be a convex optimization problem,and the optimal solution of the resource allocation policy is found using the Lagrange multiplier method.Finally,experiments are conducted using real user data sets,and the results show that the proposed method can optimize the decision policy and reduce the user response delay and backbone traffic in the long run.(2)Profit-driven joint optimization method of multidimensional resource allocation and task scheduling under dynamic computing powerThe computing-aware network supports third-party computing resource sharing.In this scenario,the uncertainty of third-party computing power and the limited computational capacity will lead to low utilization of computing resources and high task latency.Therefore,the paper proposes a profit-driven joint optimization method for multidimensional resource allocation and task scheduling to improve operator profits by enhancing resource utilization and reducing task latency.First,a profitdriven operational mechanism connecting operators,providers,and consumers is designed to model the long-term problem.Second,based on Lyapunov optimization,the long-term problem is decomposed into a series of single-time slot objectives,and long-term control is achieved by optimizing the single-time slot objectives.Then,the single time slot problem is decomposed into three sub-problems for the coupling of the single time slot problem,and a three-step decision scheme is designed.The subproblems are solved based on genetic algorithm and heuristic algorithm.Finally,experiments are conducted with real user access data sets,and the results show that the algorithm is able to limit backbone traffic and maximize operators’ profits from a long-term perspective without requiring future information.