Energy-efficient congestion control

Lingwen Gan, Anwar Walid, Steven H. LowSigmetrics 2012 · DOI

TL;DR

This paper introduces a novel dynamic bandwidth adjustment (DBA) mechanism that works alongside existing TCP/RED congestion control to save network energy by reducing link bandwidth during low traffic periods. The authors prove that their approach maintains throughput and network stability, validated through simulations.

What Problem Does This Paper Address?

The increasing energy consumption of core networks is a significant concern. While many network devices are designed to be load-independent in terms of power consumption, there's a substantial opportunity to save energy by dynamically adjusting link bandwidth, especially in core networks where link utilization can be as low as 30-40%. However, simply reducing bandwidth can lead to reduced throughput and network instability, particularly when interacting with established congestion control mechanisms like TCP and Random Early Discard (RED). This paper aims to bridge this gap by proposing and analyzing an energy-efficient congestion control mechanism that integrates dynamic bandwidth adjustment with TCP/RED, ensuring both energy savings and robust network performance.

Methods

The core of the paper is the proposed Dynamic Bandwidth Adjustment (DBA) mechanism. This mechanism operates at each link and dynamically adjusts its bandwidth based on local buffer information. Specifically, it uses an integral control law to match the link's average buffer size with a "target buffer size." This target buffer size is designed to be a function of the current bandwidth, ensuring that as bandwidth decreases, the target queueing delay also decreases, thus preserving quality of service.

The authors analyze the interaction of DBA with TCP/RED using two models:

  • Synchronous Model (SM): This model assumes ideal conditions where network delays are negligible and all components (TCP, RED, DBA) update simultaneously.
  • Simplified Synchronous Model (SSM): A simplified version of SM used for theoretical analysis, particularly for proving stability and deriving the network surplus maximization problem.
  • Simplified Asynchronous Model (SAM): This model incorporates network delays, making it more realistic for analyzing local stability.

They also propose both continuous and discrete implementations of DBA. Continuous DBA allows for fine-grained bandwidth adjustments, while discrete DBA accounts for the practical reality of bandwidth being available in fixed levels.

Datasets

Not discussed in the paper. The authors use simulations with a custom network topology.

Key Insights and Ideas

  • Energy Savings Without Throughput Sacrifice: The most significant insight is that by carefully integrating dynamic bandwidth adjustment with congestion control, it's possible to reduce link bandwidth during low-traffic periods without negatively impacting end-user throughput. This is achieved by ensuring that bandwidth reduction only occurs when there's "excess" bandwidth, and the target buffer size is dynamically adjusted to prevent packet loss.
  • Local Control for Global Stability: The DBA mechanism relies solely on local buffer information, making it easy to implement. Despite this local control, the authors prove global stability for the system under certain conditions (negligible network delay).
  • Network Surplus Maximization: The paper frames the problem as a "network surplus maximization" problem, where the goal is to maximize the sum of source utilities minus link costs. This provides a theoretical foundation for understanding the steady-state behavior of the system.
  • Trade-offs in Discrete DBA: While continuous DBA offers theoretical advantages, discrete DBA is more practical. The authors show that discrete DBA closely follows continuous DBA but exhibits some oscillation around the optimal bandwidth. This oscillation is managed by a hysteresis mechanism, prioritizing traffic protection.
  • Impact of Network Delay: The analysis highlights that network delay can challenge stability. However, the authors demonstrate that by slowing down link dynamics (e.g., increasing the update period), local stability can still be maintained even with bounded delays.

Key Results

  • Global Stability: The proposed TCP/RED/DBA system is proven to be globally stable when network delay is negligible.
  • Local Stability: Local stability is achieved when network delay is bounded and link dynamics are sufficiently slow.
  • Throughput Preservation: Simulations confirm that DBA effectively turns off excess bandwidth without reducing throughput, even in the presence of non-TCP traffic.
  • Steady-State Optimization: The steady-state source rates and link bandwidths are shown to solve a network surplus maximization problem.
  • Continuous DBA Performance: Continuous DBA exhibits fast bandwidth ramp-up (making it transparent to TCP sources) and slow ramp-down, which helps preserve network performance.
  • Discrete DBA Performance: Discrete DBA closely mimics continuous DBA's performance, maintaining throughput and generally following bandwidth trends, with some controlled oscillation.

Limitations and Future Work

  • Network Delay: While local stability is proven for bounded network delays, the analysis of global stability with significant delays is not fully explored.
  • Discrete DBA Optimization: The authors mention that improving the performance of discrete DBA is an ongoing research topic, particularly concerning the probability of bandwidth updates.
  • Practical Implementation Details: The paper focuses on the theoretical framework and simulation. Further work might be needed to address practical implementation challenges, such as the overhead of bandwidth updates in real-world hardware.
  • Non-TCP Traffic: While simulations include UDP cross-traffic, the theoretical analysis primarily focuses on TCP traffic. Further investigation into the interaction with other traffic types could be beneficial.

Summary updated

BibTeX
@inproceedings{gan2012energy,
  author = {Lingwen Gan and Anwar Walid and Steven H. Low},
  title = {{Energy-efficient congestion control}},
  booktitle = {Sigmetrics},
  year = {2012},
  doi = {10.1145/2254756.2254770},
}
Figure