Adaptive Routing

What Is Adaptive Routing?

Adaptive routing is a dynamic technology that assesses network topology and adjusts routes in response to changes in traffic load. By actively identifying link congestion, adaptive routing prioritizes selecting short and uncongested paths for packet forwarding. This enhances network throughput, resilience, and reduces latency. Presently, large supercomputing centers employ a combination of adaptive routing and direct topology. In simpler terms, it aims to minimize repetition in route selection.

Why Do We Need Adaptive Routing?

Establishing a high-performance computing center involves connecting numerous compute nodes. However, expanding the cluster can lead to increased network latency and deployment expenses, falling short of meeting computing power and deployment requirements. A direct topology offers extensive accessibility and a compact network diameter. Within such a topology, the deployment of adaptive routing serves the following purposes: Under normal network conditions, it prioritizes selecting the shortest path for packet forwarding. When the shortest path experiences congestion, an uncongested non-shortest path is chosen to forward packets. This approach ensures optimal utilization of network links, enhancing bandwidth efficiency. It fulfills the need for high throughput, low latency, and cost-effectiveness while supporting large-scale networking. In simpler terms, it minimizes repetition in route selection.

Adaptive Routing in a Direct Topology

How Does Adaptive Routing Work?

In the network, each node has multiple available forwarding paths, but only the initial node (where packets enter the network) utilizes the adaptive routing algorithm to select the best forwarding path. Subsequent nodes, after the initial one, forward packets based on their routing tables without reselecting a path.

Every network node maintains three routing tables: one for the public network (containing information about the shortest path), one for a Non-min VPN instance (holding information about non-shortest paths), and one for a Mix VPN instance (including both shortest and non-shortest paths).

When packets are received at an access port, the adaptive routing algorithm determines whether to select the shortest or non-shortest path. In this case, the routing information is sought in the Mix VPN instance routing table. For packets arriving at a min sub-interface of a global or local port, the public network routing table is consulted to find the information about the shortest path. Lastly, for packets received at a non-min sub-interface of a local port, the routing table of the Non-min VPN instance is referenced to find information about the non-shortest path. To put it simply, the routing strategy differs based on the ingress point, minimizing unnecessary repetition in the process.

The packet forwarding process can be categorized into two stages:

• The ingress node scans the best path table for the most efficient packet forwarding path, determining the packets' local outgoing interface.

• A non-ingress node searches its routing table for the packets' outbound interface depending on their inbound interface.

Ingress Node Selecting the Optimal Path

Each network node retains a best path table, which holds details regarding the most efficient forwarding route based on node routing information and link congestion status. Taking the destination IP address into account, the best path table stores information about the shortest path if it is uncongested. In cases where the shortest path is congested, details about non-shortest paths destined for the IP address are stored instead. In simpler terms, the best path table adapts based on congestion status to store the optimal forwarding path information.

Non-Ingress Node Forwarding Packets

Non-ingress nodes forward packets based on the routing table maintained by each network node, where local ports feature both min and non-min sub-interfaces at Layer 3. The min sub-interface is utilized for forwarding packets along the shortest path, while the non-min sub-interface is employed for forwarding packets along the non-shortest path.

The packet forwarding guidelines for non-ingress nodes can be summarized as follows:

• When receiving packets on a global port, the node checks the public network routing table for packet forwarding.

• For packets received on a min sub-interface of a local port, the node consults the public network routing table for packet forwarding.

• When packets are received on a non-min sub-interface of a local port, the node references the routing table of the Non-min VPN instance for packet forwarding.

In simpler terms, non-ingress nodes follow specific routing rules based on the port type to determine the appropriate path for forwarding incoming packets.