Amazon Web Services has begun deploying a new data center networking architecture across its cloud infrastructure, replacing a decades-old design model with a flatter system that the company says delivers higher throughput and lower power consumption.
The architecture, known as Resilient Network Graphs (RNG), is already operating in AWS facilities in Ireland, Germany, and Spain. AWS said the design is now the standard networking approach for most newly constructed data centers and supports a large share of its cloud workloads.
Data center networking is an unsung hero compared with GPUs or AI accelerators. Yet it plays a crucial role in cloud performance, determining how efficiently information speeds among vast arrays of servers. As cloud and AI workloads get ever heavier, networking bottlenecks offer plenty of headaches.
For decades, hyperscale data centers have relied on a hierarchical structure known as a fat-tree topology. In that design, traffic travels through multiple layers of switches and routers arranged in a tree-like hierarchy. The model’s weakness: it concentrates traffic through limited pathways, creating congestion even when unused network capacity exists elsewhere.
AWS engineers pursued an alternative approach based on random graph theory, a concept that has long been explored in academic research but has been difficult to use in real world environments. Rather than organizing routers in fixed layers, the new design connects network devices through a distributed mesh of pathways that creates many possible routes between endpoints.
According to AWS, the resulting architecture increases data throughput by as much as 33% compared with conventional designs. The company also reports a 40% reduction in power consumption associated with networking equipment.
ShuffleBox and Spraypoint
The transition required solving several engineering problems that have always prevented random graph networks from being deployed at hyperscaler data facilities.
One challenge involved managing the vast volume of fiber-optic cabling inside data centers. AWS estimates its global infrastructure contains roughly 20 million kilometers of fiber. Randomized network designs create far more complex cabling patterns than traditional architectures, making deployment and maintenance difficult.
To address that issue, AWS developed a passive optical device called ShuffleBox. The hardware requires no electrical power and standardizes cable interconnections while preserving the network’s quasi-random structure. The company says the approach simplifies deployment and allows the architecture to be replicated consistently across data centers.
However, routing data through a network with thousands of possible paths presented another challenge. Traditional routing methods typically select a limited number of preferred routes. AWS instead created a protocol called Spraypoint that distributes traffic across numerous available paths before directing it toward its destination.
Together, ShuffleBox and Spraypoint enable better use of network capacity while reducing the likelihood that traffic becomes concentrated at specific points.
AWS reports that RNG requires substantially less networking hardware than previous designs. Company figures indicate a reduction of 69% in networking devices, while infrastructure costs associated with networking can fall by as much as 45%. AWS also estimates operating costs decline by roughly 27%.
