Performance Analysis And Optimization Research In Multi-Tier Web Applications Based On Soft Resource Allocation And Mixed Workload Modeling

In order to efficiently use server resources,cloud centers(e.g.,Alibaba Cloud)deploy a large number of applications,especially multi-tier web applications,such as Taobao e-commerce systems.For multi-tier web applications,the performance of these systems will have a great impact on user experience and the profit of service providers.An Amazon report shows that the system loses 1%of its revenue for every 100 millisecond delays.The traditional way to improve system performance is to increase the number of servers or to update hardware devices.However,this traditional method not only increases the cost,but also cannot effectively improve the resource utilization of the servers.Soft resources are a series of software configurations that can determine system performance.For example,threads in the Web server.The system performance can be optimized by tuning soft resource allocation.Therefore,under the condition of the limited hardware resources,how to tune the allocation of soft resources to ensure the system performance and improve the resource utilization is a huge challenge.At the same time,the workloads of different web applications in the cloud center are different.Different workloads contain different types and proportions of requests and vary greatly in different time periods.Therefore,howto perform a fine-grained analysis on mixed workloads,and then accurately predict and optimize system performance is also a huge challenge.To address the above two challenges,this paper analyzes the performance of multi-tier web applications mainly from two aspects:soft resource allocation and mixed workload modeling.The main works of this paper are summarized as follows:1?The impact of soft resource allocation on the performance of multi-tier web applications is studied.As to multi-tier web applications,we select one soft resource in the servers of each tier.Through analyzing the impact of single soft resource allocation and multiple soft resources allocation,soft resources of different tiers are interdependence.Improving system performance needs to tune soft resource allocation reasonable.2?For the problem of soft resource allocation,a method of tuning soft resource allocation based on queueing theory and simulated annealing algorithm is proposed.This method theoretically formulates the multi-tier web applications through using M/M/N queueing models and Jackson queueing network for obtaining performance indicators.Then,the simulated annealing algorithm is applied to tune the soft resource allocation under the prerequisite of guaranteeing the system performance.3?For the performance evaluation of multi-tier web applications with mixed workloads,we firstly analyze the impact of mixed workloads on the performance of multi-tier web applications.Then the mixed workloads are modeled by discrete-time Markov-modulated Poisson process to analyze the access ratio of each type of requests in a multi-tier web application,and then we also evaluate the system performance under the condition of mixed workloads through using the queueing theory.This paper deploys the RUBiS benchmark system as the experimental environment for studying the performance of multi-tier web applications.Through using an automated testing framework,we conduct a large number of experiments and collect enough experimental measurement data,including CPU utilization of servers,data packet transmission rate of servers,log records of servers and performance indicators such as throughput and average response time.By adopting an optimization method of soft resource allocation based on queueing theory and simulated annealing algorithm,we can guarantee the system performance by optimizing the soft resource allocation for different workloads,for example,ensuring that the average response time of the system is within 1 second.By using the DMMPP to model the mixed workloads,the access probability of different requests can be obtained,comparing with the actual result,the maximum error is within 2%.In terms of the evaluation of system performance(response time),by modeling the servers of multi-tier web applications using the DMMPP/M/1 queue,the model analysis results are twice as accurate as the direct use of M/M/1 queue modeling.