Resources

Highlights

Summary

This project implements a complete multi-AS network simulation in ContainerLab using FRRouting (FRR) routers, combined with a centralized network automation system running on an Alpine Linux manager node.

By analyzing traffic prediction matrices, the manager dynamically calculates the optimal distribution of outbound traffic across multiple upstream internet links, enforcing the distribution programmatically using BGP attribute manipulation (LOCAL_PREF).

Network implementation

Complete network topology

Complete network topology

The network topology consists of three primary areas:

  • Customer networks (AS 65001 & AS 65002): Represented by customer edge routers (CE1 and CE2) and Alpine test nodes, hosting subnets 1.0.0.0/24 and 2.0.0.0/24 respectively
  • Core transit network (AS 65020): Contains provider edge routers (PE1, PE2), gateway routers (GW1, GW2), and a centralized manager node, connected together via a Linux bridge for simplicity
  • Upstream providers (AS 65004 & AS 65005): Represented by upstream routers (UP1, UP2) serving as gateways to the simulated internet (advertised prefixes: 131.114.0.0/16, 192.0.2.0/24, 198.51.100.0/24, and 203.0.113.0/24)

To avoid manual errors and improve scalability, all routers are configured via a Jinja2 template (AS numbers, router IDs, BGP neighbors, …).

Centralized traffic engineering

Standard BGP routing lacks global coordination. Outbound traffic to the internet from the core AS default-routes through a single gateway (GW1) due to BGP’s last tie-breaker rule (lower router ID).

While BGP Multipath (bgp bestpath as-path multipath-relax) can balance traffic, it operates blindly without considering traffic predictions and fails to coordinate between separate PE routers.

To solve this, our centralized traffic engineering system operates as follows:

  1. The manager parses a JSON file containing destination traffic predictions from each customer, e.g.:
    {
  "C1": { "198.51.100.0/24": 150, "203.0.113.0/24": 35 },
  "C2": { "131.114.0.0/16": 20, "192.0.2.0/24": 50, "203.0.113.0/24": 40 }
}
  2. The system sums predictions by destination prefix
    • 198.51.100.0/24 \(\rightarrow\) 150
    • 203.0.113.0/24 \(\rightarrow\) 75
    • 192.0.2.0/24 \(\rightarrow\) 50
    • 131.114.0.0/16 \(\rightarrow\) 20

    Total Load: 295

  3. To distribute \(N\) destination paths across two upstream links (GW1-UP1 and GW2-UP2) while minimizing the link utilization imbalance, the problem is solved as a binary knapsack problem where:
    • \(\text{Capacity} = \sum_{i=1}^N \text{Load}_i / 2 = 147.5\)
    • \(\text{Weight}_i = \text{Load}_i\)
    • \(\text{Value}_i = \text{Load}_i\)

Using a dynamic programming algorithm with \(\mathcal{O}(N \cdot \text{Capacity})\) time and space complexity, the manager selects a subset of paths to assign to GW1 that maximizes load without exceeding capacity (145), leaving the remaining paths (150) to GW2.

Once the optimal path distribution is calculated, the manager node programmatically enforces it by connecting via SSH into GW1 and GW2 and increasing the LOCAL_PREF attribute of the assigned routes.