Unraveling the Complexity of Network Management

Theophilus Benson, Aditya Akella, David A. MaltzNSDI 2009

Introduction

Network operators have long understood that "complex" networks are inherently harder to manage and more susceptible to catastrophic outages. However, in the realm of network engineering, "complexity" has historically remained a subjective "gut feeling" rather than a measurable metric. This lack of formalization makes it nearly impossible to compare alternative network designs, predict the risks associated with manual configuration changes, or develop automated tools to simplify management.

This paper introduces a formal framework to quantify enterprise network configuration complexity. By moving beyond misleading proxies like raw lines of code, the authors provide a systematic way to predict human effort and identify hidden configuration vulnerabilities.

Quantifying Complexity: Three Models

The authors propose three distinct models to capture different facets of network complexity:

  1. Referential Complexity: This model utilizes a referential dependence graph to map how configuration "stanzas" (such as interface definitions, ACLs, or routing protocols) point to one another. By measuring the average number of reference links per router, the authors quantify the "mental context" an operator must maintain when modifying a single line of code.
  2. Router Roles: Using a modified copy-paste detection algorithm, the authors identify "archetypal" configurations. By abstracting specific parameters like IP addresses, they determine how many distinct functional roles exist in a network (e.g., border router vs. departmental switch) and the uniformity of their implementation.
  3. Inherent Complexity: This measures the complexity of the underlying policy, independent of configuration syntax. Using reachability sets—the set of all possible packets that can travel between any two points—the authors calculate reachability entropy. High entropy indicates a network riddled with "special cases" where different paths have wildly divergent access rules.

The Dataset

The study analyzes configuration files from seven real-world networks, categorized into university and enterprise environments. The authors validated their metrics through interviews with the operators of five of these networks to ensure the mathematical models aligned with actual management difficulty.

Summary of the seven analyzed network environments, including router counts and host scales.

Key Insights: Why Lines of Code (LoC) is a Lie

A primary finding of the research is that traditional metrics like total lines of configuration (LoC) are fundamentally flawed. For example, the network Univ-2 featured relatively small configuration files but was among the most complex due to intricate, interdependent routing rules. Conversely, Univ-3 had massive configuration files that were simple to manage because they were highly repetitive and uniform.

Analysis of configuration scale and composition: (a) CDF of file sizes showing the variance in LoC, and (b) the functional breakdown of configuration files across the datasets.

Further insights include:

  • The "Cost" of Saving Money: In Univ-1, high complexity was a deliberate trade-off for lower hardware costs. Operators used a complex mixture of OSPF and RIP protocols specifically to avoid purchasing expensive OSPF licenses for certain switches, effectively trading human management effort for capital expenditure savings.
  • Dangling Configurations: A significant portion of referential complexity stems from "ghost" configurations—commands that were once functional but became irrelevant as the network evolved. Operators often leave these intact out of fear, lacking the tools to see what might break if they are deleted.
  • Predictive Power: The metrics directly correlated with the number of manual steps required for standard tasks. In the simple Enet-1, adding a subnet required only one step; in the complex Univ-1, the same task required 4-5 steps and changes to multiple routing instances.

Results and Bug Discovery

The framework's utility was proven when the "Inherent Complexity" analysis uncovered a real-world configuration bug in Univ-1. The tool detected a slight deviation in reachability entropy caused by a missing network statement in a routing stanza—an error the operators had not yet noticed.

Limitations and Future Work

While groundbreaking, the current models focus on static configuration files and do not yet account for dynamic complexity arising from link failures or route flapping. Additionally, the analysis focuses on Layer-3 routing and ACLs, omitting the impact of "middleboxes" like NATs and firewalls. Future research aims to use these metrics to automatically generate the simplest possible configuration that satisfies a specific set of reachability policies.

Summary updated

BibTeX
@inproceedings{benson2009unraveling,
  author = {Theophilus Benson and Aditya Akella and David A. Maltz},
  title = {{Unraveling the Complexity of Network Management}},
  booktitle = {NSDI},
  year = {2009},
}
Figure