Scalable techniques for network control and evaluation
The Internet has been experiencing significant changes in its scale, capacity, user population, traffic volume and the variety and quantity of online applications and devices. Its continuing evolution demands scalable technologies that provide sustained support for a network with increasing scale and complexity.
In this thesis, we propose several scalable techniques that apply to network control and evaluation. Specifically, we consider the following three problems: First, what is the appropriate congestion control algorithm for networks as capacities get higher and higher and workloads evolve? Second, how does one efficiently measure one-way packet loss rates along paths in a large network? And third, how does one simulate large high bandwidth IP networks so as to obtain detailed packet level information?
We first propose a new congestion control algorithm, E-TCP, that offers stable high performance under increasing Internet capacity and traffic volume. We develop an evolutionary network model that focuses on networks with increasing per connection throughput, increasing link capacity, and fixed router buffers. Through analysis, we illustrate the interaction between TCP congestion control, router buffer size, link capacity and link utilization. The analysis reveals a serious performance degradation problem for existing congestion control algorithms. Motivated by the analytical results, we develop a new end-to-end congestion controller, E-TCP (evolutionary TCP), that allows buffer sizes to remain constant for high link utilization and increasing data throughput.
Next, we propose a novel technique for measuring one-way path loss rates for paths in large IP networks. We use IP tunneling techniques to control the paths followed by measurement traffic in the network. Coupled with standard measurement capabilities such as NetFlow and SNMP, specific probing methods are developed to isolate the performance of groups or even individual measurement packets. We then exploit and extend multicast tomographic inference methods (MINC) to extract the performance of probe traffic on paths within the network. This yields a method that accurately determines one-way path performance with a few monitors for paths within a large IP network.
Finally, we focus on how to simulate the behavior of selected flows in an evolving network as it increases in scale and traffic. We present a hybrid simulation method that maintains the performance advantage of fluid models while providing detailed packet level information for selected traffic flows. We propose two models to account for the interaction between background TCP traffic in a fluid network and foreground packet traffic of interest. The first assumes that the packet traffic adds a negligible load on the fluid network whereas the second accounts for the added load by transforming the packet traffic into fluid flows and solving the resulting enhanced fluid model. This new technique provides us an efficient mechanism to simulate large high traffic volume IP networks as well as provide detailed packet level information for particular flows under observation.