| With the development of Big Data,Artificial Intelligence and Internet of Things,the data processing capacity of each network terminal device is greatly improved,which makes the shortcomings of the embedded system gradually appear.On the one hand,the usage scenarios of terminal equipment are relatively solidified,which leads to low utilization rate of equipment resources and excessive computing resources.On the other hand,users who need to process large amount of data information can not update high performance equipments due to cost constraint.In view of the above problems,this paper conducts the following research:This paper independently designs "CPU-FPGA-GPU heterogeneous cloud computing system"(hereinafter referred to as the CFG heterogeneous cloud computing system).The CPU in the system is used for logic control,the FPGA and GPU is applied in medical,financial and other fields requiring high data throughput and low delay.In this sysyem,users can apply for computing resources on demand and this system can dynamically load computing resources in a task-driven way,effectively avoiding excessive resources.At the same time,the hardware computing equipment is rented for the uses in this system,which can effectively avoid the huge cost of purchasing the hardware computing equipment.Due to the imperfect development of the Cyborg project,this paper does not directly use the Cyborg project to manage high-performance computing devices.This paper solves the FPGA and GPU discovery,the driver mounting,resource virtualization,interaction with Nova and the hardware resource pooling after systematically studies FPGA and GPU hardware design principles,Openstack logical architecture and Openstack virtualization.In order to solve the problems of low resource utilization and high task execution time,this paper establishes a multi-constrained two-objective task scheduling model by considering the characteristics of "CPU-FPGA-GPU heterogeneous cloud computing system" resources and tasks,the latest execution time,the system load balancing,the inter-task communications and so on.Aiming to improve resource utilization rate and enhance the task processing efficiency of CFG heterogeneous cloud computing systems,this paper proposes an up-down search algorithm for task preprocessing and an improved genetic algorithm based on group for task scheduling.The up-down search algorithm can filter out tasks that fit to run on FPGA and prepare for task scheduling.The improved genetic algorithm based on group proposes that it can place interdependent sub-tasks on the same equipment as far as possible in order to avoid the huge overhead of inter-tasks communicating across resource devices.At the same time,the improved genetic algorithm based on group adopts elite selection strategy when selecting operations to avoid elite individuals being destroyed by crossover and mutation.In the process of crossover and mutation,the constraints are added to ensure the availability of task scheduling results.Finally,two groups of experiments were carried out.Firstly,this paper simulates users to send tests to CFG heterogeneous cloud computing system and the experiment proves that this system has the characteristics of friendly user interface,flexible expansion,on-demand deployment,resource pooling and so on.Secondly,this paper evaluates the improved genetic algorithm based on group by simulation and contrast experiment.According to the results,the resource utilization of FPGA based on the improved genetic algorithm based on group is 5.4% higher than that of traditional genetic algorithm and 22.8% higher than that of random algorithm.The task execution time based on the improved genetic algorithm based on group is 4.6% lower than that of genetic algorithm and 20% lower than that of random algorithm.The results show that the improved genetic algorithm based on group can effectively improve the resource utilization of FPGA,shorten the task execution time and increase number of users. |
PDF Full Text Request