Research And Application Of Task Scheduling Strategy In Heterogeneous Surveillance Video Cloud Computing Platform

As the society pays more attention on security,the number of surveillance cameras is growing rapidly,and the video analysis task is more intelligent and complicated.While,the traditional central processing unit no longer satisfies people's need.The GPU(Graph Processing Unit)becomes a trend.It is increasingly being applied to general-purpose computing and shows high-performance much better than the traditional central processing unit.However,without consideration of the heterogeneous environment,the current cloud computing framework cannot schedule the tasks according to nodes'capability.As for the current CPU-GPU hybrid system,it lacks consideration of both the characteristic of task and the issue of dynamic allocation,it will lead to undermine the whole performance of system.Facing such challenges,this paper firstly presents a distributed task scheduling strategy for heterogeneous computing.With consideration of differences of computing speed,we use an evaluation model to unify each node's capacity.In this way,scheduling the task according to node's capacity could reduce the finish time in distributed system.Secondly,after analyzing the characteristics of video analysis tasks and the features of heterogeneity,this paper proposes a heterogeneous node task scheduling strategy based on reinforcement learning algorithm.Considering the states of computing resources keeps changing,this strategy is able to schedule tasks in run-time and adaptively select a task from task pool.By utilizing both CPU and GPU computing resources effectively,it can speed up the video processing application.Finally,in order to verify those task strategies proposed in this paper,vehicle object detection application based on Spark framework was implemented.Experimental results show that online reinforcement learning model based task scheduling strategy and distributed task strategy can significantly reduce the total completion time of the distributed surveillance video processing tasks.