This guide provides a detailed introduction to the standardized networking solution, configuration guidance, and maintenance manual for small-scale AI computing backend fabric. The solution implements a single-tier Clos network using Asterfusion data center switches, based on Rail-only architecture.

Intended for solution planners, designers, and on-site implementation engineers who are familiar with:

The Rail-only architecture is the ideal design for small-scale AI backend fabric.

As shown in the figure above, the Rail-only architecture adopts a single-tier network design, physically partitioning the entire cluster network into 8 independent rails. Communication between GPUs of different nodes is intra-rail, achieving single-hop connectivity.

Compared to the traditional Clos architecture, the Rail-only architecture eliminates the Spine layer. By reducing network tiers, it saves on the number of switches and optical modules, thereby reducing hardware costs. It is a low-cost, high-performance network architecture specifically tailored for large AI model training in small-scale compute clusters.

This example illustrates an AI cluster consisting of 32 compute nodes (128 GPUs total, 4 per server), with 4 CX732Q-N switches deployed as Leaf nodes. The key design principles are summarized as follows:

The Gateway VLAN IP address planning is as follows:

Table 1: Gateway VLAN IP Address Planning

Table 2: Configuration Overview

When connecting 400G NICs to CX864E-N switches, split each 800G port into two 400G interfaces.

Table 3: Interface Breakout Configuration

After completing the configuration, verify the interface status using the show interface summary command.

Table 4: Configuring VLAN and Interface IP Addresses

Verify VLAN configuration using the show vlan summary command.

The CX-N series switches support queues 0-7 (8 queues in total). Queue 3 and queue 4 are lossless (supporting up to two lossless queues), while others are lossy.

The default template uses system-default DSCP mapping. PFC and ECN are enabled for queue 3 and queue 4, and Strict Priority (SP) scheduling is set for queues 6 and 7.

When creating a template, you can specify three parameters:

If the provided lossless RoCE configuration does not fully suit your scenario, refer to RoCE Parameter Adjustment/Optimization for fine-tuning.

Table 5: Enabling Easy RoCE

Verify RoCE configuration using the show qos roce command.

When default configurations are insufficient, use the following commands to optimize performance.

Table 6: Modifying DSCP Mapping

If the interface has been bound to a lossless RoCE policy, unbind it before modifying.

Table 7: Modifying Queue Scheduling Policy

ECN thresholds are adjusted via min_th, max_th, and probability:

PFC thresholds are adjusted via the dynamic threshold coefficient dynamic_th:

Other parameters can remain unchanged during modification.

Recommended values for CX864E-N:

Recommended values for other models:

Table 8: Adjusting PFC and ECN Thresholds

Table 9: Interface Status Information

Table 10: Common Table Entries

Table 11: RoCE Statistics

Leaf1

Leaf2

Leaf3

Leaf4