Research On Multi-task Scheduling Strategy Based On CPU-GPU Heterogeneous System

In recent years,the homogeneous multicore processor has been unable to meet the demand of computing power.With the further development of processor structure,heterogeneous multicore processor including CPU and GPU has gradually replaced the homogeneous multicore processor as the current mainstream processor.CPU /GPU heterogeneous system has many different types of processing cores,which enables the system to process different types of tasks more efficiently,and the addition of GPU greatly improves the parallel processing ability of the system,making the running speed of the program greatly improved.CPU-GPU heterogeneous architecture not only improves the computing power of the system,but also faces many challenges.In different heterogeneous systems,it is difficult to determine the optimal matching relationship between different types of tasks and heterogeneous processing cores,and it is difficult to ensure the workload balance by task allocation between cores with significantly different processing capabilities.These problems make the resource utilization rate of heterogeneous systems low and fail to play its due performance,and the key method to solve the problem is to design reasonable System resource scheduling strategy.In order to improve the performance of CPU / GPU heterogeneous system,this paper studies the following aspects:First,considering that the thread organization structure of GPU will affect the utilization of computing resources,a block size adjustment strategy is proposed to adapt to the current application and GPU hardware environment.The strategy consists of two stages: the first stage records the execution efficiency of different thread organization structure while processing a small amount of task data by linear exploration;the second stage directly uses the most appropriate thread organization structure to process the remaining data by querying the existing records,so as to improve the efficiency of GPU.Secondly,in order to improve the resource utilization of CPU / GPU heterogeneous system,a workload balancing strategy with priority to protect the efficient core is proposed.The strategy is divided into two parts.The first part uses sample preprocessing to understand the execution efficiency of the application on the current heterogeneous core.In the second part,priority is given to the implementationof the application on the efficient core.By setting the protection threshold,we can control the reasonable allocation of tasks among heterogeneous cores,so as to ensure that the strategy will not cause negative impact in special circumstances.Furthermore,the BS-WB combination strategy can be formed by combining and reconstructing the block size adjustment strategy and workload balancing strategy,which absorbs the advantages of the two strategies and expands the application scope of the strategy.Third,in order to solve the problem that the BS-WB combination strategy is not suitable for dealing with multiple tasks,based on the idea of global task allocation,a multitask scheduling strategy is proposed to deal with a large number of different types of tasks.The strategy is divided into two parts.Firstly,the characteristics of tasks in the task group are determined according to the processing of sample data.Then,according to the information from sample processing,the efficiency of the core is given priority to assign tasks to the core for execution,and the additional cost caused by obtaining the task allocation scheme is reduced through the improved ant colony algorithm.In this paper,eight representative applications are run in CPU-GPU heterogeneous environment to test the effect of different strategies.The experimental results show that block size adjustment and workload balancing strategies can significantly improve the performance.For the application that is most suitable to enable these strategies in the selected application,these two strategies can reduce the execution time by 46.35% and 78.84% respectively.For the BS-WB combination strategy,the execution time of the eight applications is reduced by 29.13% on average.The multitask scheduling strategy reduces the total execution time by 23.38% based on the combined strategy.The combination of the two strategies improves the generality of resource scheduling strategies in heterogeneous systems,and can make better use of the processing capacity of heterogeneous systems.