deBruijn Blueprint

[Monjistha99]

Summary

A new network fabric/blueprint based on deBruijn graph topology where hosts are connected to network switches and switches are interconnected in deBruijn fashion.

• The blueprints accept switch device, server device and order.
• Size of this topology depends on order and degree.
• Degree is internally calculated where degree = switch_port.count/8.
• Total number of nodes of the graph is degree ** order

Hybrid deBruijn fabric

• A DeBruijn fabric with access layer of rack switches like Clos fabric topology.
• The fabric creates one fabric switch and one access switch per DeBruijn node, using passed switch device.
• Hosts connect to access switches, and each access switch uplinks to its corresponding fabric switch.
• Fabric link connectivity:
  0..d-1 primary outgoing
  d..2d-1 redundant outgoing
  2d..3d-1 primary incoming
  3d..4d-1 redundant incoming
  where d = degree of graph = switch_port.count/8
  remaining fabric switch ports are used for access uplinks.

Multi host deBruijn fabric

• A DeBruijn fabric where multiple hosts attach directly to each fabric switch.
• Half of the switch ports are reserved for fabric links and half for host links.
• The fabric half is split into incoming/outgoing ranges, and each range is split again into primary/redundant links

Validation

• Has switch radix validation to ensure the port count is divisible by 8.
• Adds host NIC validation to ensure the available host-facing ports. (host-facing ports // server NIC count)

Tests

• test_hybrid_deBruijn_fabric for HybridDeBruijnFabric
• test_multi_host_deBruijn_fabric for DeBruijnFabricWithMultiHost

DGX parameterized coverage using supported DGX profiles:

dgx1, dgx2, dgx_a100, dgx_h100, dgx_gb200

[Monjistha99]

added my comments wherever there was a review comment

[Monjistha99]

replied on reviews