What Is Non-Blocking Fabric?_
In a non-blocking fabric, the aggregate bandwidth of the switching infrastructure equals or exceeds the sum of all endpoint bandwidths, ensuring no single flow is forced to wait for another. This is typically achieved through a Clos or fat-tree topology with sufficient uplink capacity between leaf and spine switches. For GPU clusters, a non-blocking fabric is critical to avoid communication bottlenecks during distributed training, where all-to-all traffic patterns are common.
Technical Details
A non-blocking fabric requires that the number of uplinks from each leaf switch to the spine layer is at least equal to the number of downlinks to endpoints, ensuring a 1:1 oversubscription ratio. In practice, this is achieved by using multiple spine switches and equal-cost multipath (ECMP) routing to distribute traffic. For NVLink-based GPU-to-GPU communication within a rack, the direct point-to-point connections or NVSwitch fabric are typically designed to be non-blocking. For scale-out networks (InfiniBand or Ethernet), non-blocking design is validated by running all-to-all bandwidth tests and checking for zero packet loss under full load.
How Leviathan Systems Works with Non-Blocking Fabric
During rack assembly and GPU networking commissioning, we verify that the leaf-to-spine cabling follows the non-blocking design specified in the customer's architecture. If we find a missing uplink or a mis-cabled port, we flag it immediately because it would create a blocking point that degrades training performance.
Related service
Network Testing & Commissioning →