What is Intelligent Routing
Intelligent Routing is a load balancing approach for AI compute networks that replaces blind hash-based path selection with deliberate, traffic-aware forwarding. It is designed for the two-tier Spine-Leaf fabrics that connect GPU servers, where the goal is high-bandwidth, lossless interconnectivity between servers. (In AsterNOS, the feature is documented in the Intelligent Lossless Network Configuration Guide.)
The problem it addresses is common in AI fabrics with a 1:1 uplink-to-downlink bandwidth ratio and many ingress and egress paths. Ideally, traffic with the same destination arriving from many ingress ports would be spread evenly across all egress ports, forwarded flow by flow so packets stay in order. In practice, traffic from multiple ingress ports often converges onto a single egress port while other egress ports carry nothing. The result is uneven load sharing and congestion, even though plenty of capacity is available elsewhere in the fabric.
Intelligent Routing offers two modes to solve this. Static intelligent routing uses source-IP-based VRF assignment and policy routing to dispatch traffic deterministically. Dynamic intelligent routing senses real-time path quality and adjusts forwarding weights accordingly. Leaf and Spine devices in the same fabric must use the same mode.
How Intelligent Routing Works
Static intelligent routing
Static intelligent routing allocates a VRF to traffic based on its source IP address, isolates services through those VRFs, and forwards traffic through policy routing. Because this removes the traditional hash step from path selection, it avoids the uneven distribution that hash collisions can cause.
Route learning. After a tenant comes online, the local Leaf generates host routes for the directly connected server NICs (using ARP-to-host), places them in the appropriate tenant VRF, and advertises them with a tag. The Spine receives the tagged host routes, creates normal routes, and synchronizes them to the peer Leaf. The peer Leaf imports the tagged routes into the matching tenant VRF and attaches the information needed for policy-route forwarding.
Forwarding. When traffic arrives at a Leaf, the Leaf assigns a VRF according to the source IP and looks up the destination in that VRF's routing table. If the destination is local, it is forwarded straight to the destination NIC. If the destination is on another Leaf, the Leaf matches the policy route for that source and VRF, redirects the traffic to the specified next hop, and sends it out the corresponding uplink to a Spine. The Spine forwards it using standard Layer 3 logic, and the destination Leaf delivers it to the NIC.
Failure handling. If an uplink fails, the policy route tied to that link becomes invalid and the affected traffic falls back to ECMP across the remaining Spines, while all other traffic continues along its policy routes. Remote Leaf devices notice that the number of next hops has dropped below the number of ports in the uplink list, so their policy routes for the affected routes are also replaced by ECMP. When the link recovers, the policy routes are restored. A Spine failure is handled the same way for all Leafs, and because traffic temporarily forwards using standard ECMP during the failure, the documentation recommends combining static mode with enhanced hash mechanisms.
Dynamic intelligent routing
To understand dynamic mode, it helps to look at the three mainstream load balancing techniques it improves on:
Flow-based ECMP: the most widely used approach, hashing the 5-tuple to choose a path. It never reorders packets and works well with many flows, but suffers hash collisions in flow-sparse environments such as AI training, leading to suboptimal distribution.
Flowlet-based subflow balancing: splits flows using an inter-flowlet time gap. Configuring that gap accurately is impractical when the network has no global view of path latency.
Packet-based ECMP: theoretically the finest balancing granularity, but in practice it causes severe packet reordering at the receiver.
Dynamic intelligent routing is a sensing-driven alternative. Switches in the network measure path quality and dynamically adjust local path selection. It enhances flow-based ECMP with Dynamic WCMP (Weighted Cost Multipath), and because data centers and carriers predominantly use BGP, it carries path quality through BGP by defining a new extended community attribute.
Path quality. The assessment combines three factors observed in AI clusters: bandwidth utilization, queue occupancy, and forwarding latency. Bandwidth and queue utilization are read from ASIC hardware registers with hundred-millisecond accuracy and announced over BGP at one-second intervals, using weighted averaging that favors recent data to limit control-plane overhead. Forwarding latency is measured with HDC (High Delay Capture), an In-Band Network Telemetry technology that captures packets exceeding a user-defined latency threshold, extracting the first 150 bytes together with metadata such as ingress and egress ports and latency, and sending them to a collector.
Propagation. Path quality travels in the Path Bandwidth Extended Community attribute. When NIC2 advertises its IP to its Leaf, that Leaf passes it to the Spine along with link quality (the quality toward the NIC multiplied by the Leaf's downlink weight). The Spine advertises it to the remote Leaf, adding its own link quality and the accumulated path metric. The remote Leaf aggregates the result and generates routing instructions.
WCMP. WCMP forwards traffic across paths in proportion to their weights, with ECMP as the special case of equal weights. In the documentation's example, a path rated 38 and another rated 80 receive a roughly 3:7 weight split. As traffic and quality change, updated weights propagate over BGP and each Leaf regenerates its dynamic WCMP routes.
Failure handling. If an uplink or Spine fails, the affected Leaf withdraws the corresponding routes, other Leafs detect the withdrawal, and traffic continues over WCMP across the remaining Spines. When the link or Spine recovers, routes are re-advertised and full-path WCMP is restored.
Why Intelligent Routing is Beneficial
Uses the whole fabric: Instead of letting traffic converge on one egress while others sit idle, it distributes traffic across all available paths, reducing congestion in AI backend networks.
Avoids hash collisions in AI traffic: AI training generates few, large flows, which is exactly where hash-based ECMP performs poorly. Static mode removes the hash step, and dynamic mode adds quality awareness.
Preserves packet order: Both modes forward flow by flow, avoiding the reordering that per-packet balancing causes at the receiver.
Adapts to real conditions: Dynamic mode weighs bandwidth, queue occupancy, and latency, and shifts traffic as conditions change, rather than relying on fixed assumptions.
Built on BGP: Because path quality rides on BGP extensions, it integrates with the routing protocol that data centers already run.
Isolates services: VRF assignment by source IP keeps tenants and services separate while balancing their traffic.
Degrades gracefully: When a link or Spine fails, traffic falls back to ECMP (static mode) or continues over WCMP on the remaining paths (dynamic mode), and returns to normal operation after recovery.