Data Modelling for the Evaluation of Virtualized Network Functions Resource Allocation Algorithms

Windhya Rankothge, Franck Le, Alessandra Russo, Jorge LoboarXiv 2017 · OpenReview

Abstract

To conduct a more realistic evaluation on Virtualized Network Functions resource allocation algorithms, researches needed data on: (1) potential NFs chains (policies), (2) traffic flows passing through these NFs chains, (3) how the dynamic traffic changes affect the NFs (scale out/in) and (4) different data center architectures for the NFC. However, there are no publicly available real data sets on NF chains and traffic that pass through NF chains. Therefore we have used data from previous empirical analyses and made some assumptions to derive the required data to evaluate resource allocation algorithms for VNFs. We developed four programs to model the gathered data and generate the required data. All gathered data and data modelling programs are publicly available at github repository.

Introduction

In the domain of Network Function Virtualization (NFV), researchers face a significant bottleneck: the lack of public, realistic datasets for testing resource allocation algorithms. These algorithms are designed to optimize the placement and resource provisioning (e.g., CPU, memory) for virtualized middleboxes such as firewalls, load balancers, and Intrusion Detection Systems (IDS).

Evaluating these algorithms effectively requires comprehensive data on network policies (how functions are chained), traffic flow volumes, temporal traffic fluctuations that trigger scaling events, and the underlying physical data center topology. Because real-world data from ISPs and large enterprises is proprietary, researchers often rely on oversimplified assumptions. This paper addresses this gap by providing a suite of four data modeling programs that synthesize empirically-grounded network environments.

Methodology

The authors developed four specialized generation programs to create high-fidelity synthetic data based on existing empirical studies:

  1. Policy Request Generation (C++): This program generates service chains of 2 to 7 network functions. It utilizes a truncated power-law distribution to mirror the complexity of real-world enterprise environments.
  2. Initial Traffic Distribution (C++): Based on HTTP traffic breakdown data from an Italian ISP, this tool distributes traffic across generated policies, assuming an initial equal distribution across active chains.
  3. Scaling Requirements (C++): This models 24-hour traffic dynamics. It identifies potential bottlenecks within a chain and calculates "scale out" (adding instances) or "scale in" (removing instances) requirements based on a traffic threshold ($L$).
  4. Topology Generation (Python): This program automates the creation of data center "maps" using three industry-standard architectures: k-fat tree, BCube, and VL2.
Standard data center topologies supported by the generation tools: K-Fat Tree, BCube, and VL2.

Empirical Foundations and Insights

The modeling tools are built upon several key empirical benchmarks to ensure realism. A critical component is the modeling of enterprise-scale deployments. For instance, a "large-scale" enterprise (10k-100k hosts) typically requires approximately 100 network functions, including specific counts for IP firewalls, IDS/IPS units, and load balancers.

Empirical data on middlebox counts across different enterprise sizes used to calibrate the modeling tools.

The scaling logic adopts a "Stratos" approach, simulating a trial-and-error process to locate specific bottlenecks rather than scaling every function in a chain. This reflects more efficient, realistic operational behavior. Furthermore, the model incorporates "flash event" statistics (e.g., from the 2006 World Cup) and data center traffic magnitude changes to simulate realistic temporal dynamics.

Results and Public Availability

The primary contribution is a reproducible pipeline for NFV research. By making these modeling programs publicly available on GitHub, the authors enable the community to generate standardized testbeds. The model successfully accounts for gradual traffic changes in 15-minute windows, allowing for the simulation of complex phenomena like "elephant flows" and sudden "flash events" across various physical constraints.

Limitations and Future Work

While providing a robust framework, the model includes certain simplifying assumptions. Traffic flow is assumed to be strictly proportionate to an NF's capacity, whereas in reality, processing-to-traffic ratios vary significantly between different types of NFs (e.g., deep packet inspection vs. simple firewalls). Additionally, the current focus is primarily on HTTP traffic; future work could expand to include more diverse traffic types like VoIP or encrypted streaming, which exhibit distinct scaling behaviors.

Summary updated

Abstract

To conduct a more realistic evaluation on Virtualized Network Functions resource allocation algorithms, researches needed data on: (1) potential NFs chains (policies), (2) traffic flows passing through these NFs chains, (3) how the dynamic traffic changes affect the NFs (scale out/in) and (4) different data center architectures for the NFC. However, there are no publicly available real data sets on NF chains and traffic that pass through NF chains. Therefore we have used data from previous empirical analyses and made some assumptions to derive the required data to evaluate resource allocation algorithms for VNFs. We developed four programs to model the gathered data and generate the required data. All gathered data and data modelling programs are publicly available at github repository.
BibTeX
@misc{rankothge2017data,
  author = {Windhya Rankothge and Franck Le and Alessandra Russo and Jorge Lobo},
  title = {{Data Modelling for the Evaluation of Virtualized Network Functions Resource Allocation Algorithms}},
  year = {2017},
  eprint = {1702.00369},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1702.00369},
}

Loading related content...

Figure