What Is Adaptive Routing?_
Unlike static routing, which uses fixed paths, adaptive routing continuously monitors link utilization and latency to select the best available path for each packet or flow. This reduces hotspots and improves overall throughput in high-radix topologies common in GPU clusters. It is typically implemented at the switch level, using algorithms that balance load without requiring host-side changes.
Technical Details
Adaptive routing relies on per-packet or per-flow decisions made by switch ASICs, often using mechanisms like dynamic load balancing or congestion-aware hashing. It can react to transient congestion in microseconds, rerouting traffic around failed or oversubscribed links. In InfiniBand fabrics, this is often realized through the Adaptive Routing (AR) feature of the switch controller, while Ethernet fabrics may use similar techniques like per-packet spraying or load-aware ECMP. The exact algorithm and threshold behavior are defined by the switch vendor and must be tuned per the OEM spec for the specific fabric topology.
How Leviathan Systems Works with Adaptive Routing
During GPU cluster commissioning, we verify adaptive routing is enabled and correctly configured on all fabric switches, as it directly impacts training job stability and performance. In the field, we often test this by injecting traffic and monitoring for balanced link utilization across the spine-leaf topology.
Related service
Network Testing & Commissioning →