TCP Vegas

Paper Review: TCP Vegas: End to End Congestion Avoidance on a Global Internet

This paper introduces a new implementation of TCP, called Vegas, which
improves throughput and packet loss over the previous implementation
of TCP, called Reno, by modifying the sender while adhering to the TCP
specification.

TCP Vegas is claimed to provide the following improvement:

- 37% to 71% better throughput than TCP Reno
- Retransmits one-half to one-fifth the amount of data as TCP Reno

These performance enhancements are a result of the following three new
techniques used by TCP Vegas. First, it checks for timeouts when
receiving ACKs rather than waiting for the coarse-grained timer
(500ms). This results in a reduction of coarse-grained timeouts by a
half, leading to faster retransmission of lost packets. Secondly, TCP
Vegas employs a proactive, rather than reactive, congestion avoidance
scheme. It attempts to find the optimal operating point, without
passing it and creating packet loss. This is accomplished by
comparing actual and estimated values for throughput (using RTT and
window sizes) and adjusting the window size accordingly. Finally,
Vegas modified the slow-start mechanism to exponentially grow every
other RTT. The RTTs in between are used to detect when to enter the
linear increase mode. This prevents the common losses that occur by
continually doubling the window until packet loss.

In their experiments, Vegas seemed to easily outperform Reno in both
throughput and packet loss. Furthermore, Vegas traffic is able to
coincide with Reno traffic without starving it, thus Vegas? gains are
not at the cost of other TCP traffic. This bodes well for adaptation
of TCP Vegas. TCP Vegas was also considered more stable and resilient
to parametric changes in the network when compared to TCP Reno.
Outside of the throughput and packet loss measures, TCP Vegas was also
shown to be at least as fair as TCP Reno. On the topic of added
latency, the paper doesn?t address the issue completely. They only
mention of an improvement in latency for the tcplib TELNET test case.
Further work in this area is definitely needed.

The paper is well-written and thorough. The simulation and
experimental results validate the claims of improved throughput and
packet loss. The experiments also cover a wide range of cases and all
quantitative results are explained logically. Their scheme adheres to
the TCP specification while making minor modifications for performance
improvement. The graphs used to illustrate the results were somewhat
unclear even though care was taken to describe them. The use of more
standard graph types perhaps would improve the clarity of the analysis.

-- Nadeem Abji
Received on Wed Oct 04 2006 - 23:04:19 EDT