Thanks to a new network architecture, networks can be designed to be 40 percent more energy-efficient.
AWS is the first company to implement a flat data center network at scale, inspired by organizational charts. The Resilient Network Graphs design improves reliability and performance while reducing energy consumption by forty percent and lowering costs.
Seshadhri Comandur, a mathematician associated with AWS and the University of California, responded to a Slack message from an Amazon colleague regarding routing expertise. Together with Giacomo Bernardi, they worked on a flat network design inspired by a decades-forgotten academic Penrose Tiles theory.
The result is a breakthrough intended to improve reliability and performance for AWS customers, potentially saving billions on hardware and reducing CO2 emissions.
From theory to practice
Classic data center networks use hierarchical structures, which can lead to bottlenecks. Bernardi was convinced that a flat but structured setup could better distribute the load and eliminate single points of failure. The breakthrough came with the application of random graph theory, where routers are connected directly to each other instead of through hierarchical layers.
The team designed the ShuffleBox, a device that deterministically shuffles connections to create random graphs without physical chaos. Additionally, they developed the Spraypoint protocol, which sends data via multiple paths simultaneously to reduce congestion.
Efficiency and scalability
In tests, the new design performed up to a third faster than traditional hierarchical structures, which Amazon says delivers performance improvements. AWS expects to reduce energy consumption for network equipment by forty percent, lowering CO2 emissions in several regions. The network is currently being rolled out in Spain and Germany, with a global implementation planned for 2026.