Lawrence Berkeley National Laboratory

Most distributed applications today manage to utilize only a small percentage of the needed and available network bandwidth. Often application developers are not aware of the potential bandwidth of the network, and therefore do not know what to expect. Even when application developers are aware of the specifications of the machines and network links, they have few resources that can help determine why the expected performance was not achieved. What is needed is a ubiquitous and easy-to-use service that provides reliable, accurate, secure, and timely estimates of dynamic network properties. This service will help advise applications on how to make use of the network's increasing bandwidth and capabilities for traffic shaping and engineering. When fully implemented, this service will make building currently unrealizable levels of network awareness into distributed applications a relatively mundane task. For example, a remote data visualization application could choose between sending a wireframe, a pre-rendered image, or a 3-D representation, based on forecasts of CPU availability and power, compression options, and available bandwidth. The same service will provide on-demand performance information so that applications can compare predicted with actual results, and allow detailed queries about the end-to-end path for application and network tuning and debugging.