Research On DPDK-based Parallel Computing Scheduling Algorithm

With the rise and development of computer technology,the computing capacity of computers is also continuously improved,which promotes the further development of other fields.As an important way to improve the computing performance,availability and reliability of computer systems,parallel computing is one of the hotspots and main directions of the development of computer technology today.This article does a lot of research on the background and basic knowledge of parallel computing.To unify the model,a method for transforming independent tasks into related tasks is proposed.And make in-depth research on the existing associated task scheduling algorithms,analyzes their shortcomings.Based on the HEFT(Heterogeneous Earliest Finish Time)algorithm,the IRE(Improved Rank and EFT)algorithm is proposed.It improves calculation method for rank and EFT(Earliest Finish Time)values,to fully consider the heterogeneity of computing nodes and the impact of tasks on subsequent tasks.And It makes the scheduling length of the DAG(Directed Acyclic Graph)task graph shortened.From the operating system level,the reasons for the high communication cost are analyzed,and the mTCP user mode protocol stack is used instead of the kernel protocol stack of the operating system.And use the SIMD(Single Instruction Multiple Data)instruction set supported by modern x86 architecture CPUs(Central Processing Unit)to optimize a large number of four operations.Use setting CPU affinity and huge page memory to eliminate the performance consumption of task scheduling and switching,improve the efficiency of memory access,allow applications to occupy memory as much as possible,and reduce the occurrence of TLB(Translation Lookaside Buffer)misses.Finally,with the help of the DAG generator and SLR(Schedule Length Ratio)indicators,the effects of the number of tasks,CCR(Communication to computation ratio),and heterogeneous factors on the execution results of the algorithm were tested.And compare the software and hardware tuning to no optimization.Finally,it is proved that the IRE parallel computing scheduling algorithm proposed in this paper and its tuning have significantly improved both the scheduling result and execution efficiency.In the best case,the SLR improvement of the scheduling result can reach 27%.