Improving Network Performance through Measurement Based Analysis
MetadataShow full item record
The unprecedented increase in the scale of established networks such as the Internet and Cellular Networks and evolution of new kinds of networks such as Wireless Mesh Networks, Sensor Networks, etc., has led to several challenging research problems in maintaining and improving network performance.In this dissertation I study some of the challenges in the design and evaluation of routing protocols and examine how performance problems of existing routing algorithms can be understood and solved. I focus on how routing affects performance in two networks. the Content Distribution Networks in the Internet and Wireless Mesh Networks, which represent widely different design considerations. In both networks, there are many factors that affect the network performance. For the Internet, the routing policies, network management, traffic pattern, network dynamics are all important influential factors. In wireless mesh networks, the wireless channel characteristics, interference, environmental changes can also affect the system performance. Thus theoretical analysis considering all these factors are prohibitively difficult. I take a measurement based approach and through extensive measurements and observations derive effective mechanisms to improve their real world performance.To understand and analyze how routing affects the performance of Content Distribution Networks in the Internet we have developed a tool, called WhyHigh, to diagnose the cause for inflated latency on any given path. We have used WhyHigh to diagnose several instances of inflated latencies, and our efforts over the course of a year have significantly helped improve client performance of a large CDN. Accurate location information for nodes in the network is important in diagnosing inflated paths, hence we also examine causes of errors in localization algorithms and propose heuristics for improving the quality of localization.For wireless mesh networks, we have developed a Channel Characteristics-Aware Routing Protocol (CARP) that use measurement based analysis to model and exploit characteristics of wireless channels. We identify better paths even when the channel experiences high temporal fluctuation. We have empirically quantified the performance gain of different protocol mechanisms on a multi-hop wireless test-bed called MiNT-m.In summary, as networks scale and the networked system becomes extremely complex, theoretical models are often limited in applicability due to the large number of input parameters in a real system. Measurement based analysis, diagnosis and protocol development are complementary to theoretical analysis. This dissertation demonstrate the effectiveness of measurement based approaches in both wired and wireless network settings.