Study On Scheduling Algorithm Of Cloud Computing

As the increasing kinds of applications, and the increasing size of data, traditional resource organization and management strategy can not meet the user requirements for computing and storage. As a result, it prompts the emergence and rapid development of "Cloud Computing". Cloud Computing can be considered to a large-scale cluster, which is composed of hundreds of the machines, and provide users Computing power, data storage and applications as a service, allowing users to use resource as the user of hydropower resources. It achieves widespread concern from academia and industry. As more and more applications move to Cloud Computing platform, it is greatly significant to enhance the performance of Cloud Computing.With the development of Cloud Computing, duo to different applications require-ments, different distributed computing frameworks are born, corresponding cloud plat-forms produce. Mainstream cloud computing platforms include:(1) MapReduce cloud platform, which is suitable for offl-ine batch processing applications; (2) the event-driven cloud platform, which efficiently processes iterative and incremental applica-tions; (3) elastic cloud platform, which supports running multiple different computing framework applications at the same time.In this dissertation, for the purpose to improve the performance of Cloud Comput-ing platforms, we focus on task scheduling problem of mainstream Cloud Computing platforms, after an in-depth study of the existing Cloud Computing key technologies and research status.In this dissertation, the main works and contributions are as follows:1. First of all, we study the characteristics of the MapReduce cloud platform, and then propose a prefetching based a scheduling algorithm, and designed a task scheduler called HPSO (High Performance Scheduling Optimizer). The basics idea is to combine the task scheduler, prediction, and prefetching mechanism to-gether, and make data prefetching decision under the guidance of task scheduling strategy to increase the proportion of tasks with data locality. Therefore, HPSO can optimize the performance of cloud platform. The results show that HPSO can obtain at least 90% of the data locally, and 1.49 times speedup.2. We analysis the characteristics of the event-driven cloud platform, and proposed a two-step scheduling algorithm called TSS(Two-Step Scheduling). The first step is to directly assign the new triggers to the node where the relevant datasets set. As a result, it can quickly respond to data update operations; The second step is to resolve the cluster load unbalancing. It adopts the distributed random load bal-ancing algorithm under the macro-control decision of master node to achieve load balancing. We configured TSS on Domino, which is an open-source event-driven cloud platform, and experiments verified the effectiveness of our algorithm.3. Finally, we research the factors that influence the elastic cloud platform. This cloud platform adopts the incremental resource mechanism to prevent starvation task, and reserve some resources for the task. It results in low resource utiliza-tion. Address this problem, we proposed a new classification scheduling algo-rithm-CategoryS. The algorithm defines two labels:small job and large job. Our algorithm can be divided into two parts. The first part takes precedence large job to minimize the number of reserved resources; and in the second part, CategoryS lends the reserved resources of large job to small job. Once the waiting resources of large job arrive, loan reserve resources will be recovered. Thus our algorithm improves resource utilization and does not affect the operation of large job. We configured CategoryS on YARN, which is an open-source elastic cloud platform, and the results show that CategoryS has demonstrated performance.