Research Of On-Chip Mission And Resource Scheduling Based On Multi-FPGA

Field-Programmable Gate Array(FPGA)features low power consumption,high parallelism,and high computational speed.The next-generation FPGAs also feature dynamic reconfigurability to further enhance device flexibility and computational efficiency.FPGA replaces the X86 or GPU processing unit in the traditional high-performance computing platform,which can effectively solve the problems of high energy consumption and low computational efficiency of the computing platform,and thus is widely used by various computing service providers.Compared with general-purpose processor and GPU,FPGA has new attribute features,which bring new challenges to the system’s task resource management scheduling.On the one hand,multi-chip FPGAs can form larger-scale and more powerful computing units through high-speed inter-chip interconnection,but put forward higher requirements for task segmentation and reorganization;on the other hand,traditional scheduling strategies implement tasks and FPGAs.One-to-one mapping between resource modules with low resource utilization.This paper studies the task resource management scheduling algorithm from two aspects of task scheduling and resource management.The task scheduling algorithm generates an FPGA configuration sequence according to the subtask topology of the large task,and the resource management algorithm allocates the FPGA on-chip resources for each subtask according to the configuration sequence.The specific research contents are as follows:(1)In order to solve the problem of resource waste in resource management algorithm,an FPGA on-chip resource management algorithm based on dynamic task combination is proposed,which realizes the many-to-one mapping between task and FPGA internal resource module.According to the task configuration file(Bitstream)on the FPGA,the available resources and tasks on the FPGA chip are respectively abstracted into a resource rectangle model and a task rectangle model with a specific length and width.Based on the dynamic programming algorithm,aiming at maximizing the utilization of the resource rectangular model,the optimal combination strategy of the task rectangle model is generated.The system calls the combination strategy to implement the task resource mapping,which effectively improves the resource utilization and resource flexibility of the FPGA.(2)In order to solve the problem that the task scheduling algorithm on Multi-FPGA can’t handle multi-task concurrent configuration,a multi-sequence based multi-FPGA joint scheduling algorithm is proposed.For a single task,the data dependency matrix is used to represent the task topology,and the matrix decomposition is used to generate an effective task sequence.A valid task sequence for multiple task topologies,which is mixed based on the task deadline,and a configuration sequence is generated for each FPGA by assigning an FPGA number to the subtask.Maintain multiple sequences to achieve load balancing between tasks in Multi-FPGA,make full use of system resources,and reduce the execution time of global tasks.The evaluation function of the scheduling result is designed according to the resource usage,and the simulation result is optimized by the simulated annealing algorithm.Based on the resource requirements of common algorithms,the TGFF test data generator is used to generate two sets of test data to verify the designed algorithm.Firstly,the improved algorithm,resource management algorithm based on quadtree and resource management algorithm based on centralized bus are compared in terms of scheduling efficiency(SF),resource efficiency(RE)and maximum available white space biggest resource area(BRA),and other performance indicators,the results show that the dynamic task combination can effectively combine the tasks,the resource flexibility is increased by 6% on average,the resource utilization rate is increased by 7% on average;secondly,the joint scheduling and independent scheduling are compared.The FPGA system differs between the task execution time and the task resource occupancy.In the result,the execution time of the joint scheduling is reduced by 38%,and the average resource utilization is reduced by 13%.The result shows that the joint scheduling can effectively weaken the physical resource boundary.