Metrics for Assessing Changes in Flow-based Networks

TL;DR

This paper introduces new metrics and a framework for evaluating how individual data flows impact network performance, especially under fluctuating traffic conditions, using percentile analysis and a modified Shapley value approach.

Contribution

It presents a novel set of metrics, including the Utilization Score and flow influence measures, for assessing network load and flow impact, validated across various scenarios.

Findings

Metrics effectively capture network state changes caused by specific flows.

Three metrics provide broad insights while remaining easy to maintain.

The methodology offers a flexible framework for future network performance evaluation.

Abstract

This paper addresses the challenges of evaluating network performance in the presence of fluctuating traffic patterns, with a particular focus on the impact of peak data rates on network resources. We introduce a set of metrics to quantify network load and measure the impact of individual flows on the overall network state. By analyzing link and flow data through percentile values and sample distributions, and introducing the Utilization Score metric, the research provides insights into resource utilization under varying network conditions. Furthermore, we employ a modified Shapley value-based approach to measure the influence of individual flows on the network, offering a better understanding of their contribution to network performance. The paper reviews and compares 11 metrics across various network scenarios, evaluating their practical relevance for research and development. Our evaluation demonstrates that these metrics effectively capture changes in network state induced by specific flows, with three of them offering a broad range of valuable insights while remaining relatively easy to maintain. Moreover, the methodology described in this paper serves as a framework for future research, with the potential to expand and refine the set of metrics used to evaluate flow impact on network performance.