Performance measurement and prediction in packet-switched networks: techniques and applications
Robert Leston Carter · 1998
The characteristics of available network connections can have a large impact on the user-perceived performance of distributed applications. Unfortunately, current packet-switched networks like the Internet do not provide applications with information regarding path characteristics such as bottleneck link speed and available bandwidth. If measurements of these quantities were available, applications could make more informed decisions regarding resource utilization, potentially providing reduced response time. In order to provide these end-to-end measurements at the application level, this dissertation presents two tools: B sc PROBE, which provides an estimate of the uncongested bandwidth of a path; and C sc PROBE, which gives an estimate of the current congestion along a path. These two measures may be used in combination to provide the application with an estimate of available bandwidth between server and client. In the first part of this dissertation we discuss the design and implementation of the probe tools, specifically illustrating the techniques used to achieve accuracy and robustness. We present validation studies for both tools which demonstrate their accuracy in the face of actual Internet conditions, and give the results of a survey of available bandwidth to a random set of WWW servers as a sample application of our probe technique. In the second half of this dissertation, we show how our probe tools can be used to enable dynamic server selection, a new approach to choosing servers in a replicated-server environment. Under dynamic server selection, applications postpone the choice of server until just before the service is requested; we show that this allows application-level congestion avoidance. We show that dynamic server selection consistently outperforms static policies. Further, we show that simple measurement techniques can provide most of the benefits of dynamic server selection while minimizing network impact. Finally, through simulation, we evaluate the impact of our techniques when deployed in a global network such as the Internet. We show that dynamic server selection appears to have at least three positive effects when used ubiquitously in a wide-area network: (1) Response time at clients is significantly reduced when compared to static selection policies; (2) Overall network traffic volume is decreased; and (3) Network traffic is redistributed away from the network backbone onto regional and local networks.