FastLane: An Agile Congestion Signaling Mechanism for Improving Datacenter Performance

David Zats, Anand Iyer, Randy H. Katz, Ion Stoica, Amin Vahdat2013 · DOI

TL;DR

FastLane introduces a novel, lightweight mechanism for datacenter networks that allows switches to directly and rapidly signal congestion to senders. This agile signaling significantly reduces high-percentile flow completion times (over 80% improvement) by enabling faster responses to packet loss and congestion, all with minimal overhead.

What Problem Does This Paper Address?

Modern datacenters are experiencing a surge in demand for rich, interactive content, which translates to increasingly strict latency requirements. A key bottleneck in meeting these demands is the network's ability to handle congestion. Traditional transport protocols, like TCP, rely on indirect signals like timeouts and duplicate acknowledgments to detect packet loss and congestion. These methods are slow and imprecise, leading to significant delays, especially for latency-sensitive "short flows" (typically 2-20 KB). This slowness exacerbates the "tail latency" problem, where a small percentage of flows take an unacceptably long time to complete, impacting overall application performance. The paper aims to bridge this gap by enabling a much faster and more direct communication channel between the network's congestion points (switches) and the data senders.

Methods

The core of FastLane is a new congestion signaling mechanism. Instead of relying on end-hosts to infer congestion, FastLane empowers switches to proactively signal congestion. When a switch detects that a packet is about to be dropped due to congestion, it transforms that packet into a "congestion signal." This signal is then sent directly back to the original sender. To ensure these signals are delivered quickly and reliably, they are given the highest priority in the network.

Key aspects of the FastLane mechanism include:

  • Direct Signaling: Congestion signals travel directly from the congested switch to the sender, bypassing the receiver and its associated delays.
  • High Priority: Signals are treated with the highest priority to minimize queueing delays.
  • Transport-Agnostic: The mechanism is designed to work with any transport protocol, making it broadly applicable.
  • Packet Transformation: A to-be-dropped packet is repurposed by flipping its source and destination addresses, setting a specific TOS (Type of Service) bit in the IP header, and truncating it at the transport header. This transformation is designed to be efficient for switches to perform at line rate.
  • Sender Response: Senders receiving these signals can then make more informed decisions, such as retransmitting packets sooner, differentiating between packet loss and out-of-order delivery, and potentially utilizing multiple network paths more effectively.

The paper also details how TCP can be modified to leverage FastLane, including:

  • Using the TOS bit to identify notifications.
  • Reconstructing dropped packets using sequence numbers and data length.
  • Modifying TCP's congestion control to react more aggressively to these direct signals.
  • Introducing new TCP options to support multiple paths and reduce acknowledgment frequency.

Datasets

Not discussed in the paper. The evaluation relies on simulations and a testbed setup.

Key Insights and Ideas

  • The "Fastest Signal is Direct": The fundamental insight is that to achieve truly agile congestion response, the signal must originate at the point of congestion (the switch) and reach the sender with minimal delay. Bypassing the receiver is crucial.
  • Priority is Paramount: Giving congestion signals the highest priority is essential to ensure they don't themselves contribute to congestion and are delivered rapidly.
  • Transport Agnosticism is Key to Adoption: Designing FastLane to be independent of specific transport protocols significantly increases its potential for widespread adoption in diverse datacenter environments.
  • Overhead is Manageable: Despite the concern that sending extra packets might exacerbate congestion, the paper demonstrates that the overhead of FastLane signals is minimal (capped at 2% bandwidth and 5% buffer usage), and can even lead to an overall reduction in packet traffic by enabling more efficient TCP behavior.
  • Timeliness is Critical: The paper strongly emphasizes that even slight delays in signal delivery can lead to significant performance degradation, highlighting the importance of the low-latency, high-priority delivery mechanism.

Key Results

  • Significant Reduction in High-Percentile Flow Completion Times: FastLane achieved over 80% reduction in 99.9th percentile flow completion times in various scenarios.
  • Order-of-Magnitude Improvements in Failure Cases: In the presence of link failures, which create severe congestion hotspots, FastLane demonstrated improvements of up to an order of magnitude compared to TCP.
  • Minimal Overhead: The bandwidth overhead of FastLane notifications was shown to be less than 2%, and buffer usage less than 5%, even under extreme congestion.
  • Improved Performance Across Flow Sizes: While short flows saw the most dramatic improvements, longer flows also benefited, experiencing an average flow completion time reduction of 25%.
  • Effective Resource Capping: FastLane maintained significant performance benefits even when bandwidth and buffer resources for notifications were strictly capped.

Limitations and Future Work

  • Testbed Limitations: The paper notes that at very low utilization levels, the testbed's inability to prioritize flows could lead to less-than-ideal performance, a limitation that would be addressed by newer NICs with better prioritization capabilities.
  • Simulation vs. Real-World: While the simulator was validated against the testbed, some discrepancies were observed at low utilization due to the simulation's simplified model of end-host processing delays.
  • Future Work: The authors suggest that FastLane's transport-agnostic nature makes it suitable for integration with other protocols like MPTCP and pFabric, potentially unlocking further performance gains. They also highlight the need for further investigation into how transport protocols can best leverage the rich information provided by FastLane signals.

Summary updated

Abstract

The drive towards richer, more interactive content places increasingly stringent latency requirements on datacenters. A critical component of meeting these is ensuring that the network responds agilely to congestion, bounding network latency and improving high-percentile flow completion times. We propose a new approach to rapidly detecting and responding to congestion. We introduce FastLane, a congestion signaling mechanism that allows senders to respond more quickly. By delivering signals to senders with high probability and low latency, FastLane allows them to retransmit packets sooner, avoiding resource-wasting timeouts. It also enables senders to make more informed decisions by differentiating between out-of-order delivery and packet loss. We demonstrate through simulation and implementation that FastLane reduces high-percentile flow completion times by over 80% by effectively managing congestion hot-spots. These benefits come at minimal cost-FastLane consumes no more than 2% of bandwidth and 5% of buffers.
BibTeX
@misc{zats2013fastlane,
  author = {David Zats and Anand Iyer and Randy H. Katz and Ion Stoica and Amin Vahdat},
  title = {{FastLane: An Agile Congestion Signaling Mechanism for Improving Datacenter Performance}},
  year = {2013},
  doi = {10.21236/ada587047},
}
Figure