HSTCP-LP: A Protocol for Low-Priority Bulk Data Transfer in High-Speed High-RTT Networks

This work presents HSTCP-LP (High-Speed TCP Low Priority), a high-speed TCP stack whose goal is to utilize only the excess network bitrate (bandwidth) as compared to the “fair-share” of bitrate as targeted by other TCP variants. By giving a strict priority to all non-HSTCP-LP cross-traffic flows, HSTCP-LP enables a simple two-class prioritization without any support from the network. It enables large file backups to proceed without impeding ongoing traffic, a functionality that would otherwise require a multi-priority or separate network. One class of applications for HSTCP-LP is low-priority background file transfer over high-speed networks. Examples are bulk data transfers of huge scientific data across the Internet, database replication, or Internet content distribution. A second class of applications is available bitrate optimization (e.g., to select a mirror server with the highest available bitrate). Current techniques first estimate the available bitrate and then download data via a transport protocol. HSTCP-LP, since it only uses excess/available bitrate, is able to estimate available bitrate while doing a useful data transfer. We develop HSTCP-LP by merging two existing protocols: the first is High Speed TCP [1]; and the second is TCP-LP [2]. The goal is for HSTCP-LP to inherit the desired functionality of both, TCPLP’s ability to give strict priority to the cross-traffic, and HSTCP’s efficiency in utilizing the excess network bitrate. Moreover, since HSTCP maintains strict fairness with current (non-high-speed) TCP implementations on low-speed links [1], it consequently enables HSTCP-LP to achieve a strict low-priority service in a broad span of networking environments: vs. current TCP implementations (e.g., TCP Reno) on low-speed links (in heavy or moderate packet drop ranges), and vs. high-speed TCP implementations (e.g., HSTCP [1], Scalable TCP [3], FAST TCP [4], BI-TCP [5], and H-TCP[6]) in high-rate networks. On one hand, HSTCP-LP inherits two low-priority mechanisms from TCP-LP. First, in order to provide non-intrusive low-priority service, HSTCP-LP flows must detect oncoming congestion prior to cross-traffic flows. Consequently, HSTCP-LP uses inferences of one-way packet delays as early indications of network congestion rather than packet losses as used by the TCP-Reno-like cross-traffic flows. Second, HSTCP-LP inherits TCP-LP’s congestion avoidance policy with two objectives: (1) quickly back off in the presence of congestion from the background flows and (2) achieve fairness among HSTCP-LP flows. On the other hand, HSTCP-LP inherits HSTCP’s increase/decrease policy for large window sizes that enables it to quickly utilize and retain the available excess bitrate in the absence of sufficient cross-traffic. In summary, HSTCP-LP is a TCP-LP version with HSTCP-like agile properties, or alternatively, a HSTCP stack with built-in TCP-LP-like low-priority mechanisms. However, HSTCP-LP is far from being a trivial fusion of the two ancestor TCP stacks. The key challenge in designing the protocol is overcoming a magnified tradeoff (when compared to lower-rate links) between the ability to successfully utilize the excess bitrate on one hand and to quickly backoff in moments of congestion on the other. For example, the original TCP-LP backoff policy that radically reduces window size when detecting persistent congestion (see reference [2] for details) is not entirely applicable to a highspeed environment since it can significantly degrade HSTCP-LP’s performance. Consequently, HSTCP-LP applies a hybrid congestion avoidance scheme that utilizes TCP-LP-like mechanisms only in low excessbitrate ranges, and then converges toward the less-backoff-responsive HSTCP policy as the window size increases. SLAC-PUB-10456 May 2004