Research On Tail Latency Measurement Method Of Cloud Platform

In recent years,with the development of cloud computing,more and more Internetcompanies have chosen to deploy their own online services and offline services on public cloud.However,with more and more types of loads are deployed on the data center public cloud,the defects of the cloud platform itself are gradually leaking out.The long tail latency phenomenon is the major problem,which has plagued cloud service providers for a long time.Because of this,international mainstream cloud service providers have to limit the utilization of hardware resources of their own cloud platforms.Taking the CPU as an example,due to the long tail latency phenomenon,the utilization rate of CPU resources of cloud platforms of international mainstream cloud service providers is generally maintained at less than 20%.Such a low utilization rate has caused a huge waste of resources.People have proposed many new architectures of public cloud,from virtual cloud to zoned cloud to software-defined shared cloud.In order to evaluate the tail latency of different architectures of public cloud.The difference in performance requires people to constantly evaluate the performance of tail latency in the new cloud platform architecture.However,due to problem of efficiency in the traditional measurement methods,the cost of measurement is too high to evaluate frequently.That means the enterprise community urgently needs a new tail latency measurement method.Taking the tail latency measurement method of the cloud as the research object,this paper focuses on one main topic: evaluating tail latency in a short time.The main contributions of the paper are summarized as follows.This paper proposes the idea of improving the traditional tail latency measurement method.We can see that there are several problems in the traditional tail latency measurement process from past,such as the cost of measurement and the difficulty of evaluation.We here summarize these problems as one,that is,the inconvenience caused by the long testing time.In this paper,based on the new theoretical research results,the testing time is greatly reduced,the overall cost and measurement difficulty of the tail latency measurement of the cloud platform are reduced.And we pointed out the problems that the traditional methods may encounter by the shorter testing time.Then this paper proposes ideas for improving existing assessment methods.This paper proposes a new design ideas of test link.The method of compressing the test link time reduces the measurement cost,but the new tail latency measurement method may be too single due to the shorter test link time compared with the traditional tail latencymeasurement method,which would lead to the complexity problem in the actual production environment.In the situation,the new tail latency assessment method is difficult to comprehensively evaluate the tail latency performance of the cloud platform under various conditions.This paper puts forward a new design idea of test experiment,comprehensively considers the tail latency performance of different aspects of the cloud platform described by different test conditions.And then we predict the performance of these experimental results in terms of accuracy and stability from a theoretical perspective.This paper presents a probabilistic tail latency measurement method to public cloud.Based on the latest theory of cloud platform tail latency,this paper analyzes the reasons of the long tail latency phenomenon and proposes a probabilistic tail latency measurement method of cloud platform.And then this paper explains the correctness of this method from the perspective of algorithm logic.In order to verify the correctness of this method,this paper also proposes a verification method for the evaluation results of this evaluation method.The last part is the experimental part.Based on the improved experimental idea,this paper carries out a simulation experiment.The experimental results are used as an example to characterize the tail latency performance of the test platform.Finally,based on the previous verification methods,the evaluation methods proposed its accurate and stability.