The Study Of The Data Placement And Scheduling On Multi-Port Domain Wall Memory In Multi-Core System

Big data,artificial intelligence,cloud computing,edge computing and other technologies are developing towards the embedded direction,and the embedded systems have been widely used in the internet of things(Io T)and the industrial internet of things(IIo T).This puts forward more requirements for the memory of embedded system.In the future,embedded system should meet the requirements of time efficiency and energy consumption efficiency.Domain wall memory(DWM)has the advantages of low access latency,low energy consumption,low cost,high density,and nonvolatility.However,data access on DWM always requires moving operations to align the domain with the read/write port for read/write operations,which will generate additional overhead.Therefore,the data placement and scheduling strategy in the DWM will greatly affect the total execution time and the system performance.In order to solve the problem of data placement and scheduling in DWM-based multi-core system(DPSDMS),an integer linear progamming(ILP)strategy is proposed to obtain the optimal solution.Meanwhile,a DADG-based layering and grouping algorithm(DLGA)is proposed to obtain the approximate optimal solution in polynomial time,which DADG is the data access dependency graph.The main work of this paper includes:1.Define the data placement and scheduling in DWM-based multi-core system(DPSDMS)and give a proof of NP completeness.Firstly,the data access dependency graph and the initial conditions of multi-core system(ICMS)are difined;the DPSDMS problem is to obtain the data placement strategy and the data scheduling strategy under the conditions of a given ICMS and satisfying the data access dependency graph,reducing the data movement operations and increasing the parallelism of instruction execution in the multi-core system,so that the total execution time of executing instructions in multi-core system is shortest;finally,the NP completeness proof of the DPSDMS is given through mathematical analysis.2.The integer linear programming strategy is proposed to obtain the optimal solution.In this paper,integer linear programming strategy is used to optimize the mathematical modeling for DPSDMS,and the mathematical formulas of the definition and linear constraints are given.And get the optimal solution by traversing all data placement and scheduling conditions.However,because the execution time of this strategy increases factorially with the increase of the amount of data,the optimal solution cannot be found in polynomial time.3.The DADG-based layering and grouping algorithm strategy is proposed.If all the current layer data of the data access dependency graph is accessed each time,then the next layer is accessed,which greatly simplifies the complexity of the problem.At the same time,due to the current layer data access has a relatively high degree of relevance,and good parallelism and less movement operations can be obtained by considering only the current layer.Therefore,this paper proposes the DADG-based layering and grouping algorithm strategy to obtain the approximate optimal solution in polynomial time.This paper designs and measures the total execution time of the multi-core system based on domain wall memory.The experimental results show that compared to sequential data placement and scheduling strategy,the ILP strategy reduces the total execution time by 60.6% on average,and the DLGA strategy reduces the total execution time by 60.33% on average.