Research On Task Scheduling Technology For Dispersed Computing

In the past decade or so,cloud computing has dominated as the primary computing paradigm in providing on-demand,location-independent,and latency-tolerant services.With the rise of edge and fog computing,some computing tasks is also being offloaded to the edge of the network.But as the promotion of the 5G and the Internet of Things,massive amounts of data are generated at the edge or in the network,and more and more applications(e.g.,virtual reality,autonomous driving,etc.)begin to seek low-latency or realtime computing response,requiring location-aware computing services.The cloud computing and fog/edge computing lack effective sensing and utilization of the in-network overflow computing resources,and they cannot meet the corresponding demands of applications.Thus dispersed computing is proposed as a complementary paradigm,which can further expand the utilization of network computing and communication resources,and can provide lower latency guarantees and higher computational scalability.Task scheduling is one of the key technologies of dispersed computing.However,traditional heuristic task scheduling algorithms cannot adapt well to the high dynamics,high heterogeneity and disperse of computing environment.Meanwhile these algorithms lack the learning capability to effectively sense the dispersed computing resources.In this paper,we focus on the dynamics,heterogeneity and disperse problems of the task scheduling in dispersed computing,and aim to realize a self-adaptive task scheduling with learning capability for dispersed computing,thus to take advantage of the ”local” computing power of dispersed computing.In this paper,we propose a scalable dispersed computing task scheduling model,which models the task scheduling decision process in dispersed computing as a Markov Decision Process,and propose a Q-Learning-based in-domain task scheduling algorithm and a Double DQN-based cross-domain task scheduling algorithm for dispersed computing.The simulation experimental results show the feasibility and effectiveness of the proposed algorithms.The main contributions of this paper are listed as follows.1)We model the dispersed computing task scheduling decision process as a Markov Decision Process and establish a novel scalable dispersed computing task scheduling model.Based on the division of dispersed computing domains,the task scheduling problem in dispersed computing is deconstructed into in-domain task scheduling and crossdomain task scheduling,which reduces the problem scale of dispersed computing task scheduling;2)To address the resource dynamics and heterogeneity problems of in-domain dispersed computing task scheduling,we propose a Q-Learning-based in-domain dispersed computing task scheduling algorithm to achieve lower latency and adaptiveness.The simulation results on randomly generated task graph data show that the proposed indomain dispersed computing task scheduling algorithm outperforms the baseline algorithm in terms of comprehensive performance and can effectively reduce makespan and task execution delay.3)To address the node dispersion problem faced by cross-domain disperesed computing task scheduling,a cross-domain dispersed computing task scheduling algorithm based on Double DQN is proposed.A state space description containing resource information is designed,and a softmax-based sampling of cross-domain dispersed nodes is implemented to reduce the cluster for cross-domain dispersed computing task scheduling.The simulation results show that the proposed cross-domain dispersed computing task scheduling algorithm generally outperforms the baseline algorithm and achieves resource-aware and global scheduling optimization of cross-domain dispersed computing task scheduling.