Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PASTRAMI: Performance Assessment of SofTware Routers Addressing Measurement Instability

Paolo Lungaroni, Andrea Mayer, Stefano Salsano, Pierpaolo Loreti, Lorenzo Bracciale

TL;DR

PASTRAMI is presented, a methodology designed to assess the stability of software routers, which is critical to accurately evaluate performance metrics such as the Partial Drop Rate at 0.5% (PDR@0.5%).

Abstract

Virtualized environments offer a flexible and scalable platform for evaluating network performance, but they can introduce significant variability that complicates accurate measurement. This paper presents PASTRAMI, a methodology designed to assess the stability of software routers, which is critical to accurately evaluate performance metrics such as the Partial Drop Rate at 0.5% (PDR@0.5%). While PDR@0.5% is a key metric to assess packet processing capabilities of a software router, its reliable evaluation depends on consistent router performance with minimal measurement variability. Our research reveals that different Linux versions exhibit distinct behaviors, with some demonstrating non-negligible packet loss even at low loads and high variability in loss measurements, rendering them unsuitable for accurate performance assessments. This paper proposes a systematic approach to differentiate between stable and unstable environments, offering practical guidance on selecting suitable configurations for robust networking performance evaluations in virtualized environments.

PASTRAMI: Performance Assessment of SofTware Routers Addressing Measurement Instability

TL;DR

PASTRAMI is presented, a methodology designed to assess the stability of software routers, which is critical to accurately evaluate performance metrics such as the Partial Drop Rate at 0.5% (PDR@0.5%).

Abstract

Virtualized environments offer a flexible and scalable platform for evaluating network performance, but they can introduce significant variability that complicates accurate measurement. This paper presents PASTRAMI, a methodology designed to assess the stability of software routers, which is critical to accurately evaluate performance metrics such as the Partial Drop Rate at 0.5% (PDR@0.5%). While PDR@0.5% is a key metric to assess packet processing capabilities of a software router, its reliable evaluation depends on consistent router performance with minimal measurement variability. Our research reveals that different Linux versions exhibit distinct behaviors, with some demonstrating non-negligible packet loss even at low loads and high variability in loss measurements, rendering them unsuitable for accurate performance assessments. This paper proposes a systematic approach to differentiate between stable and unstable environments, offering practical guidance on selecting suitable configurations for robust networking performance evaluations in virtualized environments.

Paper Structure

This paper contains 15 sections, 14 equations, 122 figures, 1 table.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Reference configurations
  4. Methodological aspects
  5. Metrics for load-loss and load-delivery analysis
  6. Ideal vs. real load-loss curves
  7. Evaluating the representativeness of PDR@0.5%
  8. Assessment Methodology
  9. Identification of saturation conditions
  10. How to assess the accuracy of PDR@0.5%
  11. Detailed design and implementation
  12. Description of experiments
  13. Conclusion and Future Work
  14. Acknowledgements
  15. Appendix

Figures (122)

  • Figure 1: Reference configurations
  • Figure 2: Layer impacting performance in different example configurations
  • Figure 3: Ideal load-loss curve (saturation offered load is 1 Mpps, the saturation throughput equals the saturation offered load.
  • Figure 4: A more realistic synthetic load-loss curve (in green) along with the load-delivery curve (in red). The saturation offered load is 1 Mpps, the saturation throughput equals 0.99 of the saturation offered load (PLR at saturation is 1%), the trashing threshold $O^\text{trash}$ is 1.5 Mpps and $M$ is 1 Mpps.
  • Figure 5: Both the ideal (blue) and the more realistic (green) load-loss curves are plotted in log scale, considering as the bottom of scale a loss ratio of $10^{-6}$.
  • ...and 117 more figures