Teraflows over Gigabit WANs with UDT

The TCP transport protocol is currently inefficient for high speed data transfers over long distance networks with high bandwidth delay products (BDP). The challenge is to develop a protocol which is fast over networks with high bandwidth delay products, fair to other high volume data streams, and friendly to TCP-based flows. We describe here a UDP-based application level transport protocol named UDT (UDP-based Data Transfer) with these properties and which is designed to support distributed data-intensive computing applications. UDT can utilize high bandwidth efficiently over wide area networks with high bandwidth delay products. Unlike TCP, UDT is fair to flows independently of their round trip times (RTT). In addition, UDT is friendly to concurrent TCP flows, which means it can be deployed not only on experimental research networks but also on production networks. To ensure these properties, UDT employs a novel congestion control approach that combines rate-based and window-based control mechanisms. In this paper, we describe the congestion control algorithms used by UDT and provide some experimental results demonstrating that UDT is fast, fair, and friendly.