SwiftQueue

Per-packet latency prediction for optimizing L4S queue selection

SwiftQueue Architecture

Overview

SwiftQueue is a novel L4S (Low Latency, Low Loss, and Scalable Throughput) queue-selection strategy that uses a custom Transformer-based latency predictor to dynamically assign packets to queues on a per-packet basis, rather than the traditional per-flow approach.

The Problem

L4S is an emerging router queue management technique that assigns packets to queues based on header markings. Current L4S implementations use per-flow queue selection, meaning all packets in a flow use the same queue. This approach can harm tail latency when transient congestion affects individual packets differently.

Key challenges with existing approaches:

  • Per-flow limitation: All packets marked the same way, even when individual packets experience varying conditions
  • Tail latency issues: Transient congestion affects some packets more than others
  • Static decisions: No adaptation to real-time network conditions

What SwiftQueue Does

SwiftQueue introduces per-packet queue selection using machine learning:

  • Per-packet Latency Prediction: Uses a custom Transformer architecture to predict the latency of each outgoing packet
  • ACK Pattern Analysis: Analyzes latencies from recently received ACKs to forecast future packet behavior
  • Dynamic L4S Marking: Sender dynamically marks packet headers to route packets to appropriate queues
  • Within-flow Differentiation: Packets within the same flow can be assigned to different queues based on predicted latency

Key Features

  1. Transformer-based Prediction: Leverages the expressiveness of Transformers for sequential pattern recognition
  2. Real-time Decision Making: Predicts latency for each packet as it is sent
  3. L4S Integration: Works with existing L4S-enabled routers
  4. Latency Spike Detection: Identifies packets likely to experience latency spikes or drops

Results

  • 45-65% more accurate latency prediction compared to state-of-the-art methods
  • 36-45% reduction in tail latency for L4S-enabled flows
  • Validated using real network traces

Resources

Citation

@inproceedings{ray2026swiftqueue,
    title = "{SwiftQueue: Optimizing Low-Latency Applications with Swift Packet Queuing}",
    author = {Siddhant Ray and Xi Jiang and Jack Luo and Nick Feamster and Junchen Jiang},
    booktitle = {New Ideas in Networked Systems (NiNeS)},
    year = {2026},
    month = feb,
    address = {Online}
}