The quality of network connectivity between a pair of Internet hosts can vary greatly. Adaptive applications can cope with these differences in connectivity by choosing alternate representations of objects or streams or by downloading the objects from alternate locations. In order to effectively adapt, applications must discover the condition of the network before communicating with distant hosts. Unfortunately, the ability to predict or report the quality of connectivity is missing in today's suite of Internet services. To address this limitation, we have developed SPAND (shared passive network performance discovery), a system that facilitates the development of adaptive network applications. In each domain, applications make passive application specific measurements of the network and store them in a local centralized repository of network performance information. Other applications may retrieve this information from the repository and use the shared experiences of all hosts in a domain to predict future performance. In this way, applications can make informed decisions about adaptation choices as they communicate with distant hosts. In this paper, we describe and evaluate the SPAND architecture and implementation. We show how the architecture makes it easy to integrate new applications into our system and how the architecture has been used with specifics types of data transport. Finally, we describe LookingGlass, a WWW mirror site selection tool that uses SPAND. LookingGlass meets the conflicting goals of collecting passive network performance measurements and maintaining good client response times. In addition, LookingGlass's server selection algorithms based on application level measurements perform much better than techniques that rely on geographic location or route metrics.

We analyze the way in which Web browsers use TCP connections based on extensive traffic traces obtained from a busy Web server (the official Web server of the 1996 Atlanta Olympic games). At the time of operation, this Web server was one of the busiest on the Internet. We first describe the techniques used to gather these traces and reconstruct the behavior of the TCP on the server. We then present a detailed analysis of the TCP's loss recovery and congestion control behavior from the recorded transfers. Our two most important results are: (1) short Web transfers lead to poor loss recovery performance for TCPs, and (2) concurrent connections are overly aggressive users of the network. We then discuss techniques designed to solve these problems. To improve the data-driven loss recovery performance of short transfers, we present a new enhancement to the TCP's loss recovery. To improve the congestion control and loss recovery performance of parallel TCP connections, we present a new integrated approach to congestion control and loss recovery that works across the set of concurrent connections. Simulations and trace analysis show that our enhanced loss recovery scheme could have eliminated 25% of all timeout events, and that our integrated approach provides greater fairness and improved startup performance for concurrent connections.

The Internet is a very large scale, complex, dynamical system that is hard to model and analyze. In this paper, we develop and analyze statistical models for the observed end-to-end network performance based on extensive packet-level traces (consisting of approximately 1.5 billion packets) collected from the primary Web site for the Atlanta Summer Olympic Games in 1996. We find that observed mean throughputs for these transfers measured over 60 million complete connections vary widely as a function of end-host location and time of day, confirming that the Internet is characterized by a large degree of heterogeneity. Despite this heterogeneity, we find (using best-fit linear regression techniques) that we can express the throughput for Web transfers to most hosts as a random variable with a log-normal distribution. Then, using observed throughput as the control parameter, we attempt to quantify the spatial (statistical similarity across neighboring hosts) and temporal (persistence over time) stability of network performance. We find that Internet hosts that are close to each other often have almost identically distributed probability distributions of throughput. We also find that throughputs to individual hosts often do not change appreciably for several minutes. Overall, these results indicate that there is promise in protocol mechanisms that cache and share network characteristics both within a single host and amongst nearby hosts.