Algorithm Research On Online Load Balancing With Fixed Length Tasks In Cloud Data Centers

Cloud Computing providers deploy diverse geographical data centers based on Internet, aiming to satisfy the requirements of various users. With virtualization technology, Cloud Computing becomes more complex and large scale. Combined with heterogeneous distributed systems, cloud data centers resource management and allocation comes into a quite challenging issue. Resource scheduling is essential for Cloud data centers. Regard to resource scheduling algorithms, real-time requests and their life-cycle are not considered in most existing algorithms, or only single type of resource is considered. In this thesis, I model and design a resource scheduling algorithm under online situations, considering multi-dimensional resource and requests lifecycles. Moreover, I extend an algorithm for offline scenario.It’s difficult to research all problems under real environment. On the one hand, network environment is not controlled by developers; the other hand, network environment is too complex. Therefore, large-scale distributed systems research would be built on simulation tools. Data center simulation system defines application loads, like user information, data centers locations, user and data center number, resource in data centers and so on. In addition, the simulation system could also simulate requests generation and allocation process. With simulation system, application developers could evaluate strategies for resource allocation, choose suitable data center for specific requests and reduce costs in data centers. Although there are already some simulation tools developed, like CloudSim, which is based on previous tools and quite heavy to execute. In this thesis, FlexCloud, a novel simulation system is designed and implemented, focusing on light weight design and resource scheduling algorithms.The major contributions of this thesis are as following: 1) design and implement a flexible and extendible simulation system for data center resource scheduling, compared with the state-of-art Cloud Sim, it saves more memory and costs less running time; 2) the two proposed scheduling algorithms in this thesis overwhelm other algorithms in imbalance degree, makespan and other load balance indexes; 3) the proposed algorithms are implemented under real test-bed and algorithm efficiency and strength are validated.