As the Internet continues to become more heterogeneous, applications must adapt to network conditions to improve the user experience. There are two modes of adaptation: over long time scales (on the order of minutes) and over short ones (on the order of connection round-trip times). On long time scales, applications need performance estimates to complete tasks such as selecting between mirror sites or choosing among various file formats. On shorter time scales, applications need information about the current congestion state of the network to decide when and what data to transmit.

In this talk, Seshan will describe two systems that address these needs. The SPAND (Shared PAssive Network performance Discovery) system is designed to provide applications with predictions about the performance of transfers between hosts. SPAND passively monitors the networks in a local domain and collects performance statistics about transfers between the local domain and distant hosts. The system uses the performance statistics of past network transfers to predict future performance. The CM (Congestion Manager) system helps applications monitor and adapt to the progress of active transfers. CM achieves this by exposing an API that enables applications to learn about network characteristics, pass information to CM, and schedule data transmissions. This API provides a simple way for applications to perform correct congestion control and enables applications to adapt their transmissions to the current congestion state of the network.

The combination of SPAND and CM provides applications the necessary tools to adapt their behavior to the state of the network. Various adaptive applications have been built using these systems. In this talk, Seshan will also describe an application that uses SPAND to control the transcoding of data to match network speeds and an implementation of TCP that uses CM to perform congestion control.

Additional information about SPAND is available at http://www.cs.berkeley.edu/~stemm/spand and information about CM is available at http://wind.lcs.mit.edu/projects/cm. SPAND was developed jointly with Mark Stemm who was a graduate student at UC, Berkeley and CM was developed jointly with Prof. Hari Balakrishnan at MIT.