LINT: Accuracy-adaptive and Lightweight In-band Network Telemetry

Shihabur Rahman Chowdhury, Raouf Boutaba, Jérôme François

TL;DR

This paper introduces LINT, a novel In-band Network Telemetry (INT) mechanism that intelligently adapts telemetry data reporting in the data plane to reduce overhead without significantly sacrificing monitoring accuracy. LINT achieves this by using a lightweight predictor to estimate potential data loss at the collector, enabling devices to selectively report telemetry, leading to a ~25% reduction in data plane overhead while maintaining high recall for critical network monitoring queries.

What Problem Does This Paper Address?

Traditional network monitoring methods, like pull-based systems, offer only coarse-grained visibility. While streaming telemetry and In-band Network Telemetry (INT) provide near real-time, per-packet visibility, INT comes with a significant data plane overhead. This overhead arises from each network device on a packet's path adding telemetry data, increasing packet size and reducing network goodput. The paper addresses the critical challenge of reducing this INT data plane overhead while ensuring that the accuracy of network monitoring queries, which rely on this telemetry data, is not severely degraded.

Methods

The core of LINT is its accuracy-adaptive reporting mechanism, inspired by model-driven data acquisition techniques, specifically the Spanish Inquisition Protocol (SIP). Instead of reporting all telemetry data for every packet, LINT devices use a predictor function to estimate the potential error at the collector if a particular telemetry data item is not reported.

Here's how it works:

  • Predictor Function: For each telemetry data item (e.g., hop latency, queue occupancy), LINT maintains two Exponentially Weighted Moving Average (EWMA) values: one for observations seen by the device (dmext) and one for observations reported to the collector (dmext).
  • Selective Reporting: A device reports a telemetry data item only if the absolute difference between the predicted value at the device (dmext) and the predicted value at the collector (dmext) exceeds a predefined error threshold (δ). This means telemetry is reported only when the device estimates that the collector would significantly misinterpret the network state without that data.
  • Lightweight Implementation: LINT is designed to run on commodity programmable network devices (like PISA-based switches) with minimal computational resources, avoiding complex operations like floating-point arithmetic, multiplication, or division. EWMA can be efficiently computed using bit shifts.
  • LINT-flow Extension: To further improve accuracy, LINT-flow considers the flow context of a packet. It maintains EWMA values per flow, allowing for more context-aware reporting decisions.

Datasets

The paper evaluates LINT using publicly available real network traces:

  • Facebook's production web service cluster (FB-WS) [19]: This trace contains metadata extracted from sampled packets (e.g., packet size, source/destination racks and pods).
  • Campus data center network (UNIV1 data set) [20]: This trace provides packet captures.
  • Wide-area network traffic capture from CAIDA (Equinix 2016) [21]: This trace also provides packet captures.

The authors also mention using a 4-port fat-tree data center network topology for their evaluation, which they set up using Mininet and bmv2 switches running a P4 program.

Key Insights and Ideas

  • The "Cost of Visibility" is Real: The paper quantifies the significant data plane overhead INT introduces, showing how even a few hops can increase packet size by 10-90%, directly impacting goodput. This highlights the practical need for optimization.
  • Accuracy vs. Overhead is a Tunable Knob: The parameter δ (error threshold) in LINT acts as a direct control for this trade-off. A higher δ means more data can be dropped, reducing overhead but potentially impacting query accuracy.
  • Data Plane Intelligence is Key: LINT demonstrates that intelligent decision-making can happen directly in the data plane, without complex control plane intervention or global coordination. This is crucial for high-speed network devices.
  • Flow Context Matters: The LINT-flow extension shows that considering the flow context can significantly improve monitoring query recall, even when selective reporting is employed. This suggests that understanding traffic patterns can lead to more efficient telemetry.
  • EWMA is Sufficiently Powerful: The choice of EWMA as the predictor function, despite its simplicity, proves effective for balancing overhead reduction and accuracy, showcasing that complex models aren't always necessary for data plane operations.

Key Results

  • Overhead Reduction: LINT can reduce INT data plane overhead by approximately 25% compared to regular INT.
  • Query Accuracy (Recall): For critical queries like identifying flows with high latency (QTail) and congested switches (QCongestion), LINT achieves a recall of over 0.9 even with significant overhead reduction (e.g., when δ = 2-5). QCongestion, in particular, maintains a recall of over 0.95 even at higher δ values.
  • Query Accuracy (NRMSE): For latency and queue occupancy queries (QLatency, QQueue), LINT achieves very low Normalized Root Mean Squared Error (NRMSE), often less than 2% for latency and less than 0.25% for queue occupancy at the chosen operating point (δ = 2-5).
  • LINT-flow Improvement: The LINT-flow extension further improves recall for QTail and QCongestion, demonstrating the benefit of flow-context awareness.
  • Per-packet Overhead: While LINT's worst-case per-packet overhead can be similar to INT, its higher quartiles are significantly smaller, meaning the typical overhead is much lower.

Limitations and Future Work

  • Parameter Tuning: The performance of LINT is sensitive to the δ parameter, requiring careful tuning to balance overhead and accuracy for specific network conditions and query types.
  • Protocol Modifications: Implementing LINT requires modifications to INT protocol message formats, specifically how to communicate which telemetry data items were reported. The paper suggests using a bitmap in unused header space but notes this requires further investigation.
  • Future Work:
    • Full-fledged implementation of LINT on commercially available PISA devices.
  • Investigating the use of information theory measures for quality-of-information-aware adaptive telemetry collection.
  • Exploring more computationally demanding predictors (e.g., machine learning-based) for future exploration.
  • Further investigation into the impact of flow context and memory limitations on LINT-flow's performance.

Summary updated

Abstract

International audience
BibTeX
@inproceedings{chowdhury2021lint,
  author = {Shihabur Rahman Chowdhury and Raouf Boutaba and Jérôme François},
  title = {{LINT: Accuracy-adaptive and Lightweight In-band Network Telemetry}},
  booktitle = {IM},
  year = {2021},
}
Figure