Low-Diameter Network Topologies for Scalable Interconnects

69 concepts33 papers2 proposals

As computing systems scale toward exascale performance and million-core many-core architectures, the primary bottleneck shifts from raw processing power to the efficiency of the communication fabric. In these massive systems, the "distance" between nodes—measured by the number of intermediate hops data must traverse—directly dictates latency, power consumption, and congestion. Researchers are increasingly moving away from traditional grid-based structures toward low-diameter network topologies that ensure any two nodes remain only a few hops apart, regardless of system size.

The Cost of the Hop

In a large-scale interconnect, every "hop" through a switch or router adds a tax of latency and energy. Each intermediate node must buffer, arbitrate, and forward the packet. In traditional 2D or 3D torus and mesh networks, this cost scales poorly; as the number of nodes $N$ increases, the network diameter (the longest shortest path) grows polynomially. Davis et al. identified this early on, proposing a hexagonal mesh with a twisted torus wrap-around to minimize diameter in symbolic multiprocessing.

Modern research emphasizes that reducing diameter is not just about speed, but about resilience. Jha et al. demonstrated that low-diameter topologies like Dragonfly, when paired with adaptive routing, reduce the duration of network congestion by an order of magnitude compared to 3D tori. By providing more direct paths and higher path diversity, these networks prevent localized "hotspots" from paralyzing the entire system.

High-Radix and Algebraic Constructions

The shift toward low-diameter designs is enabled by the rise of high-radix switches—routers with a large number of ports. By increasing the "degree" of each node, architects can create denser connectivity. Rao et al. note that while higher radix increases internal switch complexity, it significantly reduces the total hop count across the network.

To arrange these connections optimally, researchers employ sophisticated algebraic and graph-theoretic constructions:

  • Slim Fly: Blach et al. describe the deployment of the Slim Fly topology, which uses algebraic graph theory to limit the maximum distance between any two switches to just two hops. This design maximizes the efficiency of switch ports, allowing for massive scaling with fewer cables and switches than traditional hierarchical "Fat Tree" designs.
  • Circulant Graphs: Huang et al. use exhaustive searches to find optimal circulant graphs that outperform hypercubes, offering significant performance gains in high-performance computing (HPC) benchmarks.
  • Kautz Graphs: For distributed systems, Li et al. show that Kautz topologies can achieve optimal diameters and remain nearly congestion-free while maintaining a constant number of connections per node.

Physical and Geometric Constraints

While low-diameter graphs are mathematically elegant, they must be physically realizable. In the world of chiplets and wafer-scale integration, the physical arrangement of components determines the feasibility of the wiring. Iff et al. introduced HexaMesh, which arranges rectangular chiplets in a hexagonal lattice to reduce communication latency by 19% compared to standard grids.

At the extreme end of integration, Ma et al. propose a switch-less Dragonfly architecture on wafers. By distributing switching logic across a fabric of chiplets, they eliminate the need for discrete, power-hungry high-radix switches while maintaining the low-diameter benefits of the Dragonfly topology.

The effectiveness of these designs is rooted in a fundamental property of graph theory: expansion. Dinitz et al. established that large graphs with small diameters are inherently good expanders. This means that a network designed for low latency (low diameter) naturally possesses the structural robustness and throughput capacity (expansion) required for high-bandwidth parallel workloads.

However, challenges remain in making these topologies "plug-and-play." Nguyen et al. highlight the need for "layout-conscious" designs like DSN-F that allow for incremental expansion without massive rewiring, a common weakness of rigid algebraic constructions. As we move toward exascale, the goal is to balance the mathematical ideal of a low-diameter graph with the physical realities of cables, ports, and power.

Key Foundational Work

The Architecture of the FAIM-1 Symbolic Multiprocessing System (1985) — contributed 2 concept elements to this direction.
Node-disjoint paths and related problems on hierarchical cubic networks (2002) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_low_diameter_network_topology_7a3d805a,
  title = {low-diameter network topology},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/low-diameter-network-topology/rdr_7a3d805aaa21482d96226befc4ab5e8e},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_low_diameter_network_topology_7a3d805a_v1,
  title = {low-diameter network topology},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/low-diameter-network-topology/rdr_7a3d805aaa21482d96226befc4ab5e8e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake