TL;DR
This paper introduces TCP Westwood (TCPW), an end-to-end TCP congestion control algorithm that improves performance, especially over wireless links, by estimating available bandwidth. Unlike TCP Reno, which aggressively reduces its congestion window upon packet loss, TCPW uses this bandwidth estimate to set a more appropriate window size, leading to faster recovery and significantly higher throughput.
What Problem Does This Paper Address?
Traditional TCP congestion control, like TCP Reno, relies on packet loss as the primary indicator of network congestion. This works reasonably well in wired networks where packet loss usually signifies a congested router. However, in wireless networks, packet loss can also be caused by unreliable radio channels (fading, interference). TCP Reno misinterprets these wireless losses as congestion, leading to unnecessary and drastic reductions in its congestion window. This severely impacts throughput and efficiency, especially in mixed wired and wireless environments where the same TCP connection might traverse both types of links. The challenge is to differentiate between congestion-induced loss and wireless channel loss to avoid over-penalizing the connection.
Methods
TCP Westwood's core innovation is an end-to-end bandwidth estimation mechanism implemented at the TCP sender. It continuously monitors the rate of incoming ACKs to estimate the current available bandwidth of the connection.
Key techniques include:
- ACK Rate Monitoring: The sender tracks the arrival times of ACKs and the amount of data they acknowledge. This rate is used as a sample of the connection's bandwidth.
- Low-Pass Filtering: To smooth out variations and noise in the ACK arrival rate (especially due to delayed ACKs or varying inter-arrival times), a discrete-time low-pass filter is applied. This filter is designed to be adaptive to time-varying inter-arrival times.
- Virtual Sample Insertion: To maintain the effectiveness of the low-pass filter even during periods of no ACK arrivals (e.g., when the sender is idle), virtual samples are inserted to ensure the sampling interval constraint is met.
- "Faster Recovery" Algorithm: Upon detecting a congestion episode (three duplicate ACKs or a timeout), TCPW uses its estimated bandwidth to set the congestion window (cwin) and slow start threshold (ssthresh). Instead of blindly halving them as in Reno, TCPW sets them to values consistent with the estimated effective bandwidth, allowing for a quicker return to higher throughput.
A significant advantage highlighted is that TCPW is a sender-side modification and does not require any changes to intermediate network devices or the receiver, adhering strictly to the end-to-end principle.
Datasets
Not discussed in the paper. The evaluation relies on simulations using the LBL network simulator ('ns' ver.2).
Key Insights and Ideas
- Bandwidth Estimation as a Differentiator: The fundamental insight is that by accurately estimating the available bandwidth, TCPW can make more informed decisions about congestion window adjustments. This allows it to distinguish between true congestion and transient wireless losses.
- End-to-End Simplicity: The fact that TCPW achieves significant improvements without requiring network-layer modifications or proxy nodes is a major win for its practical deployability and adherence to the Internet's design philosophy.
- "Faster Recovery" is Key: The mechanism of setting cwin/ssthresh based on estimated bandwidth, rather than just halving them, is crucial for rapid recovery after congestion events, especially in lossy wireless environments.
- Friendliness and Fairness: The paper demonstrates that TCPW is not only effective but also friendly to existing TCP Reno connections, avoiding starvation and maintaining a reasonable level of fairness, even when sharing links with different RTTs.
Key Results
- Significant Throughput Gains: TCPW demonstrated substantial throughput improvements across various scenarios.
- Up to 550% improvement in mixed wired and wireless networks.
- Up to 394% gain over Reno or SACK with independent (Bernoulli) errors.
- Up to 567% gain at a propagation time of 100ms.
- Up to 300% improvement in fading conditions (bad state error rate of 5%).
- 167% improvement with blackout durations of 0.1 sec.
- Improved Fairness: TCPW showed fairness at least as good as, and often better than, TCP Reno, particularly when connections had different end-to-end round-trip times (RTTs).
- Friendliness: TCPW connections did not starve TCP Reno connections. In mixed scenarios, TCPW connections gained slightly, while Reno connections experienced a minor loss, but overall coexistence was effective.
- Performance Comparable to Link-Layer Solutions: TCPW's performance was comparable to localized link-layer approaches like Snoop, but without the overhead of specialized protocols.
Limitations and Future Work
- Limitations:
- The paper acknowledges that TCPW performs poorly when random packet loss rates exceed a few percent.
- In scenarios where the bottleneck is in the backward path (receiver to sender), TCPW might face challenges in fairly sharing bandwidth between data packets and ACKs.
- The effectiveness of TCPW in scenarios with very high error rates (e.g., >5% in bad states) is limited, as all protocols struggle.
- Future Work:
- Incorporating the NewReno feature into TCPW for more efficient recovery from multiple losses within the same window.
- Developing mechanisms to address fairness issues when the bottleneck is in the backward path.
- Further investigation into TCPW's robustness against high random packet loss rates.
- Refining the bandwidth estimation and filtering methods to improve "friendliness."
- Developing control theoretical models to analyze TCPW's stability.
- Further comparison with link-layer techniques like Snoop, especially in complex wireless scenarios.
Summary updated