TL;DR
This paper identifies a "TCP starvation" problem when legacy TCP interacts with optimal cross-layer scheduling in wireless mesh networks (WMNs). They propose a simple "fake-ACK" algorithm to mitigate this issue, allowing TCP to better adapt to the WMN's optimal rates without significant performance degradation.
What Problem Does This Paper Address?
The future of the internet, particularly wireless mesh networks (WMNs), is moving towards "clean-slate" designs that aim for theoretical optimality rather than incremental improvements. A key component of these optimal designs involves cross-layer optimization (CLO), where different network layers cooperate to maximize performance. Specifically, optimal CLO in multi-hop wireless networks often relies on max-weight scheduling for link layer efficiency and a corresponding source rate control algorithm.
However, a significant challenge arises when legacy protocols, like the ubiquitous TCP, are used by end-users to access these optimally designed WMNs. The authors highlight a critical incompatibility: TCP's congestion control mechanism, which relies on acknowledgments (ACKs) to gauge network conditions and adjust its sending rate, can be severely hampered by the deterministic scheduling policies of optimal CLO. This leads to a "TCP starvation" problem, where TCP connections can fail to establish or transmit data effectively because their packets are perpetually de-prioritized by the network's optimal scheduler, preventing them from receiving the necessary ACKs to increase their sending rate.
Methods
The core of the paper's contribution lies in identifying and proposing a solution to the TCP starvation problem.
- Problem Identification: The authors analyze the interaction between TCP's congestion window (cwnd) adjustment mechanism and the max-weight scheduling policy used in optimal cross-layer algorithms. They observe that when a TCP packet arrives at a mesh AP with a low "price" (often related to queue length or queue length difference), it may not be scheduled if neighboring nodes have higher prices. This prevents the TCP packet from being forwarded, thus delaying or preventing the reception of its ACK. Without ACKs, TCP cannot increase its cwnd, leading to starvation.
-
Proposed Remedy: Fake-ACK Algorithm: To address this, they propose a "fake-ACK" algorithm implemented at the mesh AP. This algorithm introduces a two-level queue structure: a TCP adaptation queue and a data-link queue. When a TCP packet is moved from the adaptation queue to the data-link queue (ready for transmission), the mesh AP generates a "fake-ACK" packet. This fake-ACK is sent back to the TCP sender with a specific advertised window (AWND) value.
The fake-ACK mechanism aims to provide TCP with a consistent stream of acknowledgments, even if the actual data packets are delayed. This allows TCP to continue its congestion window adjustment (both slow-start and congestion avoidance phases) as if it were receiving ACKs for its data. By controlling the rate at which these fake-ACKs are generated and the AWND value, the mesh AP can indirectly influence the TCP sender's rate to better match the optimal rates dictated by the CLO algorithm.
Datasets
Not discussed in the paper. The authors mention using GloMoSim for simulations, which is a network simulator, not a dataset in the traditional sense.
Key Insights and Ideas
- TCP's Blindness to Optimal Scheduling: TCP, designed for loss-based congestion control on stable links, is fundamentally ill-equipped to handle the deterministic, queue-length-driven scheduling of optimal cross-layer algorithms. It interprets lack of ACKs as congestion, not as a scheduling delay.
- The "Price" Problem: In max-weight scheduling, nodes with longer queues (or queue differences) have higher "prices" and are prioritized. Legacy TCP, especially during connection setup (SYN packets) or when its queue is initially small, can have a low price, leading to its starvation.
- Fake-ACK as a Bridge: The fake-ACK algorithm acts as a crucial bridge, decoupling TCP's ACK-driven rate control from the actual packet delivery delays imposed by the optimal scheduler. It provides TCP with the necessary feedback to keep its cwnd growing.
- Minimal Modification: The proposed fake-ACK solution is designed to be a simple addition to the mesh AP, requiring minimal changes to existing TCP stacks or the core CLO algorithms.
Key Results
The paper presents simulation results demonstrating the effectiveness of the fake-ACK scheme.
- Performance Equivalence: Figure 5 shows that the fake-ACK scheme achieves nearly the same throughput as the ideal cross-layer source algorithm. This indicates that the fake-ACK effectively allows legacy TCP to adapt its rate to the optimal CLO rate.
- Mitigation of Starvation: Without fake-ACK, nodes further away from the destination (higher hop count) experience significant throughput degradation because their packets are consistently de-prioritized. With fake-ACK, the throughput for these nodes is significantly improved, showing that the starvation problem is mitigated. The paper states that "throughput degradations in 4 and 5 hop nodes compared to 3 hop nodes are little."
Limitations and Future Work
- Loss-Free Assumption: The paper assumes the single-hop access network (user to AP) is loss-free. The authors acknowledge that real-world wireless access networks can be lossy, and their scheme would need to be extended to handle this.
- Reliable Packet Delivery: The fake-ACK scheme focuses on TCP congestion control adaptation. It does not explicitly describe a reliable packet delivery scheme for the WMN backhaul, which would be necessary for a complete solution.
- UDP and Other Protocols: The proposed remedy is specifically for TCP. The authors note that other protocols like UDP (e.g., for VOD streams or DNS requests) might also suffer from similar starvation issues, and future work should address how to adapt them.
- Real-World Testing: The proposed scheme needs to be verified and tested in a real WMN environment, beyond simulations.
Summary updated