On Performance And Energy Optimization Of Many-core Systems Using Approximate Computing

As multimedia processing,machine learning,and other applications can tolerate error,the processors or networks-on-chip can aggressively improve performance or reduce energy consumption by reducing the computation effort(e.g.,by reducing precision or dropping packets,etc.).This emerging computing paradigm is called approximate computing.The problem of how to optimize system performance within the constraints of quality requirements and power is key,as the quality of the output results of the application affects the user experience.However,existing approximate computing approaches in many-core systems have the following problems: most of them only target a certain component(e.g.,processor,networks-on-chip,adder,multiplier,etc.).For example,approximate networks-on-chip compresses communication packets to reduce the network load,and the application reduces the number of iterations of the loop to reduce the computational workload.Because the optimization problem of approximate computing in many-core systems has many decision variables,and the search space is large,a single control variable cannot find the optimal solution.Therefore,this dissertation studies the approximate computing in many-core systems with multiple control variables.This dissertation defines a performance optimization problem of many-core systems,that is,minimizing the execution time of application under the constraints of error and power.To solve this problem,this dissertation develops the error model,performance model and power model of the application program,proposes an optimization method based on interior point algorithm,and executes the application program according to the solution obtained by this method.Because the optimization is completed before the program runs,it is called static optimization method.In the experimental evaluation,compared with the loop truncation method,ABDTR and MMBS in the literature,the proposed method can reduce the program execution time by 33.1% and the energy consumption by 36.6%.However,the above method has the following shortcomings: 1)the error variance of the output results is large and there are some outputs that are above the error threshold,which will affect the user experience.2)the impact of the input data on quality is not taken into account,which results in some input data not being adequately approximated.In order to solve these problems,a runtime performance optimization problem of many-core systems is defined:minimizing the execution time of application programs under the condition that every output result satisfies the error constraint.A dynamic control method is proposed to solve the problem.The method finds the optimal solutions of several control variables according to the system state and the quality of output elements,and updates them at runtime to optimize performance and energy consumption.The experimental results show that,compared with the static method,the dynamic method reduces application execution time by 16.93%,energy consumption by13.52%,and the number of application output results whose errors are over 20% of the error threshold by 68.9%.