Research On Performance Optimization Of Iaa S Cloud Services Toward Multi-VMs Application

Along with the rapid growth of cloud computing, a lot of applications migrate to the cloud. Meanwhile, the scale of applications and the number of users is rising fast, making it more and more necessary to provide great performance to applications. Infrastructure clouds have the advantages of low cost and high scalability. Thus, the research on optimizing application performance based on infrastructure clouds has important academic significance and practical prospects.Currently, the complicated management mechanisms and highly dynamic character of infrastructure clouds bring big challenges to optimize application performance. First, virtual machines act as the base container to encapsulate application components. On one hand, their slow creation process seriously harm the efficiency of deploying and scaling applications. On the other hand, their performance directly restrict the overall runtime performance of applications. Second, the high sharing degree of cloud resources makes it difficult to guarantee fair performance among applications. In particular, the best-effort provision manner of network resources is hard to guarantee the fairness of network performance. Such unfairness causes many performance optimization mechanisms losing efficiency. Third, the high dynamic nature of clouds increase the difficulty of evaluating applications. The traditional evaluation approaches cannot satisfy such nature. Their high error rates lower down the efficiency and accuracy of optimizing application performance.To address these issues, we research on optimizing the performance of applications consisting of multiple virtual machines in infrastructure clouds. We consider the aspects of application performance, performance fairness, and performance evaluation, studying related theories and key mechanisms about optimizing application performance deeply from multiple angles. On one hand, we extend image format and storage manner. On the other hand, we construct performance fairness model and user behavior model. With these works, we improve application initialization and runtime performance.Specifically, the contributions of this dissertation contain four aspects.1. A mechanism of fast deploying and auto scaling applications based on the incremental technique. By introducing the copy-on-write incremental technique to the process of creating virtual machines, we eliminate the tightly coupled relation between virtual machines and their images. The approach of sharing images greatly reduces the data transmission during the concurrent creation of multiple virtual machines. Thus, it significantly improves the creation efficiency of applications. We also propose an automatic scaling method with transparent framework and scalable rules, which effectively increases the timeliness and automation degree of scaling applications. We compare our mechanism with the traditional deploying and scaling mechanisms. The evaluation result shows that our mechanism can reduce the deploying time to less than 5 seconds. Meanwhile, it improves the write performance of applications by 24%-38%, and decrease the performance degradation by 13%-17%.2. A mechanism of organizing and managing application images based on the zone storage model. By introducing the cache technique to virtual machine images, we increase the scalability of image formats. A layered image storage method maximizes the functionality and performance of images. The zone storage theory is proposed to improve the cache efficiency and thus the application performance. With an approach of aggregating disks, the zone storage is more efficient and compatible to realize. We also propose a mechanism-friendly virtual machine placement policy, which greatly improves the hit rate of cache data in the zone storage. We compare our mechanism with the traditional image organizing mechanisms. The evaluation result shows that our mechanism can improve application performance by 100% in general, and by more than 10 times while adopting the friendly virtual machine placement policy. At the same time, the scalability and availability of the mechanism is also verified in multiple situations.3. A mechanism of guaranteeing fair application performance based on scheduling virtual machines. By analyzing the traditional bandwidth fairness guarantee mechanisms, we construct a fairness model for application performance with considering response time. According to the complexity of transactions inside of applications, we propose an approach to guarantee fair performance toward application transactions. It avoids the drawback of traditional approaches toward components which is inefficient for applications consisting of multiple virtual machines. Based on the fairness model, we propose a mechanism of guaranteeing fair performance among applications by aggregating virtual machines. Then, we optimize their placement policies in order to maximize the fairness degree. By comparing with some traditional mechanisms, the evaluation result shows that our mechanism can increase the fairness degree of application performance by 26.5%-52.8%. It can avoid the performance degradation at the same time.4. A mechanism of auto testing and evaluating applications based on the customer behavior model. By reconstructing the traditional customer behavior model, we build a new model that considers transaction characteristics. Based on this new model, we propose an automatic transforming approach to timely and effectively converting the traditional model to ours. Meanwhile, we design and implement a framework to automatically evaluate application performance according to the architectures of infrastructure clouds and applications. This framework can avoid the low efficiency of manually evaluation caused by the high dynamic nature of infrastructure clouds. By comparing with the traditional evaluation mechanisms, our mechanism can reduce the error rate by 50%. Furthermore, it realizes automatic evaluation process.The above mechanisms and approaches run through multiple aspects of optimizing multi-VMs application performance. They not only cover the researches on virtual machine structure, deploying policy, and scaling policy, but also cover the designs on performance fairness model and user behavior model. These works have been deeply evaluated with mainstream mechanisms in real clouds and accepted optimization mechanisms in academic. The results prove that the above mechanisms and approaches are feasible and efficient.