Using Approximate Computing To Ensure The Performance Of Multiple Latency-sensitive Programs Running Together

The Internet has changed the world with an amazing speed.The development of cloud computing and big data industries are supported by the data center.The data center plays a very important role in today's society,and has been integrated into our lives.As a carrier for a large number of programs running together,the data center has gone from ideal to reality.However,the average resources utilization rate of large data center is only between 10% and 50%.Workload consolidation is a common method to improve the resource utilization of the data center.It refers to run multiple programs on a single server node.more than 60% of data center operators have adopted workload consolidation.Otherwise,the program content of data center is diverse.According to the performance indicators,they can be divided into batch programs measured by IPC(instruction per cycle)and latency-sensitive programs measured by response time.In addition,the dramatic performance degradation for the latency-sensitive programs is unacceptable.To ensure the performance of latency-sensitive programs,meanwhile to consolidate workloads efficiently,common research was to run latency-sensitive programs together with batch programs on the same server.Current work proposed many resource allocation and program scheduling strategies to achieve a trade-off between the latency-sensitive program performance and system resource utilization.However,programs in the data center are in the transition period from batch programs to latency-sensitive programs,and the program architecture is being redesigned.Therefore,there are many shortcomings in the method of running a latency-sensitive program together with batch programs.However,there is just a begin about the research on consolidating of multiple latency-sensitive programs.The study found that approximate computing can improve the performance of programs at the cost of accuracy loss,which can guarantee the performance of latency-sensitive programs.There are many latency-sensitive programs with high fault tolerance in the data center,such as machine learning,face recognition,image processing,which belong to approximate computing programs.When the output accuracy of the program is reduced,the difference in output quality is imperceptible to the user in vision and hearing,which conforms to the characteristics of approximate computing.Therefore,latency-sensitive programs can be divided into approximate latency-sensitive programs and non-approximate latency-sensitive programs.The data center operators can negotiate with users on the output precision range of the program,and reduce some loop iteration or instruction execution in the program to keep the original execution time of the program by using the characteristics of approximate computing.Based on the study above,this paper proposes a method to ensure the performance of multiple latency-sensitive programs running together by using approximate computing.The purpose of the method is running multiple latency-sensitive programs together on the same server node through approximate computing.On the other hand,this method guarantees the performance of multiple latency-sensitive programs and achieves a trade-off between system resource utilization,the performance of multiple latency-sensitive programs,and flexible program scheduling.The contributions of this paper are as follows:An approximate calculation method based on the LLVM compiler is designed.LLVM compiler is employed to achieve the approximate computing of software-level technique loop perforation by modifying the loop-simplify of the transformation loop.The perforation rate is set to control the approximate degree,which is represented by N,and N is an integer greater than 1.And three quality management methods are set up.Experimental results show that loop perforation can reduce the execution time of the program,and there is a direct relationship between the performance rate of the program and the perforation rate.This approximate computing method can also maintain the original execution time of the program.We propose a method to ensure the performance of multiple latency-sensitive programs running together by using approximate computing.The method runs multiple latency-sensitive programs on the same server node,that is,non-approximate and approximate delay-sensitive programs can run together;Cgroups resource management mechanism is used to limit the resources that can be used by approximate latency-sensitive programs so that the non-approximately latency-sensitive programs can be executed in the sufficient resource environment to ensure their performance;The approximate latency-sensitive program can be run in limited resources,set up the dynamic adjustment mechanism,and the original execution time of the program can be maintained by the approximate computing method.The experiment shows that the system resource utilization can be improved by running the approximate and non-approximate latency-sensitive programs on the same server node.This method can also effectively solve the problem of running a latency-sensitive program with batch programs,not only guarantees the performance of multiple latency-sensitive programs but also improves the utilization rate of system resources.Through the above work,this paper aims to ensure the performance of multiple latency-sensitive programs and improve the utilization of system resources.The LLVM compiler is used to realize automatic loop perforation and reduce the execution time.By running multiple latency-sensitive programs together on the same server node and using approximate computing technology to ensure the performance of multiple latency-sensitive programs and improve system resources utilization.Finally,realize the trade-off among the utilization of system resources,the performance of multiple latency-sensitive programs,and flexible program scheduling.