Packet losses increase latency for Web users. Fast recovery is a key mechanism for TCP to recover from packet losses. In this paper, we explore some of the weaknesses of the standard algorithm described in RFC 3517 and the non-standard algorithms implemented in Linux. We find that these algorithms deviate from their intended behavior in the real world due to the combined effect of short flows, application stalls, burst losses, acknowledgment (ACK) loss and reordering, and stretch ACKs. Linux suffers from excessive congestion window reductions while RFC 3517 transmits large bursts under high losses, both of which harm the rest of the flow and increase Web latency.

Many high performance distributed applications require high network throughput but are able to achieve only a small fraction of the available bandwidth. A common cause of this problem is improperly tuned network settings. Tuning techniques, such as setting the correct TCP buffers and using parallel streams, are well known in the networking community, but outside the networking community they are infrequently applied. In this paper, we describe a tuning daemon that uses TCP instrumentation data from the Unix kernel to transparently tune TCP parameters for specified individual flows over designated paths. No modifications are required to the application, and the user does not need to understand network or TCP characteristics.

With the growth of high performance networking, a single host may have simultaneous connections that vary in bandwidth by as many as six orders of magnitude. We identify requirements for an automatically-tuning TCP to achieve maximum throughput across all connections simultaneously within the resource limits of the sender. Our auto-tuning TCP implementation makes use of several existing technologies and adds dynamically adjusting socket buffers to achieve maximum transfer rates on each connection without manual configuration.Our implementation involved slight modifications to a BSD-based socket interface and TCP stack. With these modifications, we achieved drastic improvements in performance over large bandwidth delay paths compared to the default system configuration, and significant reductions in memory usage compared to hand-tuned connections, allowing servers to support at least twice as many simultaneous connections.

We have developed a Forward Acknowledgment (FACK) congestion control algorithm which addresses many of the performance problems recently observed in the Internet. The FACK algorithm is based on first principles of congestion control and is designed to be used with the proposed TCP SACK option. By decoupling congestion control from other algorithms such as data recovery, it attains more precise control over the data flow in the network. We introduce two additional algorithms to improve the behavior in specific situations. Through simulations we compare FACK to both Reno and Reno with SACK. Finally, we consider the potential performance and impact of FACK in the Internet.