AWS Flattens Data Center Networking With Resilient Network Graphs

AWS is putting a long-running data-center networking idea into production: flatter fabrics that borrow from random graph theory instead of sending most traffic through a familiar hierarchy of aggregation and spine layers.

The fresh signal is that AWS is now discussing Resilient Network Graphs, or RNG, as real infrastructure rather than a lab result. The Register reported on June 13 that AWS says the topology can be up to a third faster and up to 40 percent more energy efficient than traditional hierarchical designs. AWS’s own research blog gives the headline numbers behind the claim: RNG is used in AWS data centers, is the default for most new builds globally, uses 69 percent fewer routers, delivers up to 33 percent better throughput, and projects a 40 percent reduction in network-equipment electricity consumption.

What AWS changed

Traditional data-center networks are easy to reason about because they look like trees. A server talks up through top-of-rack and aggregation layers, across the fabric, and back down toward the destination. That structure makes cabling and routing manageable, but it can also concentrate congestion in the wrong places while leaving other capacity underused.

RNG attacks that problem by making the graph flatter and less dependent on a small set of higher-level choke points. The Amazon-authored arXiv paper describes RNG as a quasi-random flat data-center network deployed in production, with a distributed routing protocol and a passive optical cabling device designed to make the topology practical at scale. The paper’s abstract says RNG can match or exceed fat-tree performance across traffic patterns while being up to 45 percent cheaper.

Why random had to become practical

The word “random” is the catch, because purely random cabling is not something operators want to maintain in a production data center. Earlier academic designs such as Jellyfish showed why random graphs were attractive, but also exposed operational problems: cabling complexity, harder routing state, and designs that could be awkward for normal rack-level locality.

AWS’s answer is not a fully random mess of fiber. The Amazon Science post calls the production design “quasi-random,” preserving random-graph benefits while using ShuffleBoxes to make flat-network cabling practical. That distinction matters: the value is not randomness for its own sake, but enough path diversity to avoid overspending on the hierarchy.

Spraypoint and ShuffleBox are the operational pieces

Two implementation details carry much of the story. The first is Spraypoint, AWS’s routing approach. Amazon says typical multipath routing schemes for flat topologies could require 20 to 80 times more memory than commodity hardware provides, while Spraypoint uses random “spraying” plus destination waypoints to expose many paths without that routing-state blowup.

The second is ShuffleBox, a passive optical device that scrambles fiber connections in a controlled way. Matt Rehder, AWS’s VP of global network engineering, told The Register that ShuffleBox lets AWS make a random graph “feel more structured” by plugging fibers into a device that internally randomizes where ports emerge. In plain terms, AWS is trying to get graph-theory path diversity without turning the data center into cable spaghetti.

Where RNG fits, and where it does not

The rollout is not a blanket replacement for every network AWS operates. The Register says RNG is being used for AWS’s core database-server networks, while machine-learning hardware still uses AWS’s UltraServer network because those workloads need full-bandwidth behavior. That boundary is important: cloud providers do not have one universal fabric for every traffic pattern.

AWS has reportedly deployed RNG in Ireland, Germany, and Spain, with a plan to bring it to the majority of company data centers by the end of the year. The practical reason is straightforward: if a provider can get similar or better performance with fewer routers and lower power draw, the savings show up in capital cost, electricity use, failure domains, and the pace at which capacity can be built.

Why customers should care

Most AWS customers will never choose “RNG” from a console menu. The impact is indirect: lower networking cost pressure, better east-west traffic behavior for suitable workloads, and potentially more resilient capacity inside the regions where AWS deploys the design.

The bigger lesson is also useful beyond AWS. Data-center scale is forcing infrastructure teams to revisit assumptions that were once convenient: hierarchy is easy to operate, but not always the most efficient way to spend silicon, optics, power, and floor space. RNG is a reminder that the next cloud performance improvement may come from topology and operations engineering, not only from faster chips.

Featured image: fiber optic cable photo by Bquast, dedicated to the public domain under CC0 via Wikimedia Commons. Image cropped to 16:9 and converted to WebP.