Performance Estimation Of Multithreaded System On Heterogeneous Multi-core

Emerging embedded devices face diverse multithreaded workloads along with conflicting needs of high performance and low power consumption,which necessitates the appearance of single instruction set heterogeneous multi-core processors.This architecture can reach optimum energy efficiency ratio by employing reasonable task scheduling policy,whose bases are the accurate evaluation and prediction of performance and power consumption.Therefore,it is necessary to firstly build the multithreaded performance model to benefit task scheduling.In this thesis,a dynamic heterogeneous multi-core performance estimation model for multithreaded system is proposed and instructions per cycle(IPC)is used to evaluate the proposed model.Firstly,the analytical form of the performance model is deducted to consist of instruction-level parallelism(ILP)and thread load in theory.Secondly,the single-thread ILP prediction model is built involving the factor of instruction mix,cache miss rate,branch misprediction and issued instructions per cycle in pipeline,by means of standard linear regression using the least squares method.Thirdly,a method of cross-core thread load contribution estimation is proposed,based on which the relationship between thread load contribution and real thread load is deduced to build the multithreaded load model by simulating Linux scheduling policy.Using Gem5 simulator with the CPU models of Minor(in-order)and arm_detailed(out-of-order),the proposed model is verified by SPEC CPU 2000 and a set of synthetic applications with thread load contribution from 1%to 99%as benchmarks in this thesis.The simulation results indicate that the average errors of single-thread ILP prediction model are 6.26%and 5.9%respectively on two different cores,and those of thread load contribution prediction are 0.37%and 0.39%.In case of arm_detailed CPU model,the average error of load allocation prediction is 1.64%,and that of thread load contribution deduction is 1.11%,which reduces errors by 20%-30%on medium load fragment compared with the current load model.The final multithreaded performance estimation model incurs an error of 9.34%and 0.77%respectively on two different cores with full load,while the error is 1.41%and 12.22%respectively when both of the CPUs are under loaded.