High Performance and Cost Efficient Architectures for Clos Data Center Networks
May 12, 2018
Light Engineering room 250
Advisor: Prof. Yuanyuan Yang
This dissertation focuses on high performance and cost efficient architectures for Clos data center networks. It consists of four correlated topics.
The first topic deals with virtual machine placement problems in fat-tree data centers. Upon the arrival of a cloud tenant, a set of virtual machines should be allocated. The locations of the virtual machines need to be carefully selected because it has significant impact on (1) the performance of the virtual network from the cloud tenant's point of view, (2) the utilization of the infrastructures from the cloud provider's perspective. The major challenge comes from the contradictory nature of these two objectives. We propose a set of virtual machine placement schemes to deal with this challenge.
The second topic extends the first one. After the virtual machines are allocated, network devices and links need to be selected. The virtual link establishment is also challenging because of the conflict between sufficient bandwidth provisioning and the scarcity of the network resources. To solve this problem, we design algorithms which are aware of the configurations of the hardware, and elaborately balance the workload all over the network.
The third topic, novel architectures for data center networks, arises from the previous two topics. A classical fat-tree data center network is highly regular in terms of its structure, which results in some problems. For example, the design space of fat-tree data center networks is coarse because one single parameter, the port count of each switch, completely defines the entire topology of a fat-tree. Staring from these observations, we design a more precise and flexible model for data center networks. With this model, the infrastructure of the data center could be tailored to fit the exact demands.
The fourth topic deals with theoretical problems which arises from the first three topics. In the first three topics, we frequently use theoretical results to optimize our design. It is desirable that the theoretical results themselves can be improved. Therefore, we manage to improve the bound of the cost for multistage networks in general, which provides significant contribution to the research community.