Data centers, facilities with communications network equipment and servers for data processing and/or storage, are prevalent and essential to provide a myriad of services and applications for various private, non-profit, and government systems, and they also form the foundation of cloud computing, which is transforming the technological landscape of the Internet. With rapid deployment of modern high-speed low-latency large-scale data centers, many issues have emerged in data centers, such as data center architecture design, congestion control, energy efficiency, virtual machine placement, and load balancing.
The objective of this thesis is multi-fold. First, an enhanced Quantized Congestion Notification (QCN) congestion notification algorithm, called fair QCN (FQCN), is proposed to improve rate allocation fairness of multiple flows sharing one bottleneck link in data center networks. Detailed analysis on FQCN and simulation results is provided to validate the fair share rate allocation while maintaining the queue length stability. Furthermore, the effects of congestion notification algorithms, including QCN, AF-QCN and FQCN, are investigated with respect to TCP throughput collapse. The results show that FQCN can significantly enhance TCP throughput performance, and achieve better TCP throughput than QCN and AF-QCN in a TCP Incast setting.
Second, a unified congestion detection, notification and control system for data center networks is designed to efficiently resolve network congestion in a uniform solution and to ensure convergence to statistical fairness with “no state” switches simultaneously. The architecture of the proposed system is described in detail and the FQCN algorithm is implemented in the proposed framework. The simulation results of the FQCN algorithm implemented in the proposed framework validate the robustness and efficiency of the proposed congestion control system.
Third, a two-level power optimization model, namely, Hierarchical EneRgy Optimization (HERO), is established to reduce the power consumption of data center networks by switching off network switches and links while still guaranteeing full connectivity and maximizing link utilization. The power-saving performance of the proposed HERO model is evaluated by simulations with different traffic patterns. The simulation results have shown that HERO can reduce the power consumption of data center networks effectively with reduced complexity.
Last, several heterogeneity aware dominant resource assistant heuristic algorithms, namely, dominant residual resource aware first-fit decreasing (DRR-FFD), individual DRR-FFD (iDRR-FFD) and dominant residual resource based bin fill (DRR-BinFill), are proposed for virtual machine (VM) consolidation. The proposed heuristic algorithms exploit the heterogeneity of the VMs’ requirements for different resources by capturing the differences among VMs’ demands, and the heterogeneity of the physical machines’ resource capacities by capturing the differences among physical machines’ resources. The performance of the proposed heuristic algorithms is evaluated with different classes of synthetic workloads under different VM requirement heterogeneity conditions, and the simulation results demonstrate that the proposed heuristics achieve quite similar consolidation performance as dimension-aware heuristics with almost the same computational cost as those of the single dimensional heuristics.
|
![[NJIT logo]](http://archives.njit.edu/vhlib/images/njit_logo.gif){kind=logo}
![[Van Houten Library logo]](images/vhl-top-image.jpg)
