Skip to navigation

Fabric link calculation

View as MarkdownOpen in Claude

FabricLinkCalculation calculates administrative link targets from deployment-supplied rules and discovered compute and switch positions. The calculator does not read hardware, select fabric members, or change switch configuration. V2 layout planning uses these rules from workflows.fabric_layout in RMS TOML. See Fabric layout catalog for the configuration format and validation boundaries.

Platform wiring parameters are deployment data. The calculator contains no platform-specific interface numbering or connectivity tables. The example below uses synthetic interface names and numbers.

Rule format

FabricLinkCalculation supports TOML deserialization through Figment, which RMS already uses. Call validate() after loading rules. expand() also validates the rules before calculation.

default_state = "down"
[[rules]]
prefix = "port"
base = 1
offsets = [0, 20]
width = 2
state = "up"
terms = [
{ node = "compute", coordinate = "tray_index", divisor = 1, multiplier = 2, remap = { "3" = 3, "4" = 4 } },
{ node = "switch", coordinate = "tray_index", divisor = 1, multiplier = 1 },
]

For each selected compute/switch pair, each rule emits numbered interfaces:

prefix + (base + sum(transformed_coordinate * multiplier) + offset + lane)

Coordinates are used as supplied; the calculator does not renumber positions. Each term requires node (compute or switch) and coordinate (tray_index or slot_number). The term reads that position field, divides by divisor, optionally applies modulus, optionally looks up the result in remap, then multiplies by the signed multiplier. Division and remainder use unsigned integer arithmetic. Each offset emits width consecutive interfaces, starting with lane 0.

For compute tray 3 and switch tray 2, the example emits port9, port10, port29, and port30. Slot numbers do not affect this example.

Inputs and results

The caller supplies the participating positions and a set of controlled interface names. The result contains every controlled interface for every supplied switch. Rules set emitted interfaces to their specified state; all other controlled interfaces receive default_state.

The calculator does not infer membership or validate chassis identity. The caller must supply positions from the same physical chassis and the intended compute and switch groups. Restoration targets and device readback belong to the caller. When integrating with a restoration map, derive the interface-name set from that map’s keys; the calculator does not need a second configured interface list.

Validation

Unknown fields, node kinds, coordinates, and states fail deserialization. Missing node or coordinate selectors also fail deserialization. The shared node-kind type includes powershelf, which calculator validation rejects. Rules, offsets, inventory, and controlled interfaces must be nonempty. Widths, divisors, and moduli must be positive. Remaps must be nonempty, with canonical decimal u32 keys. Prefixes contain lowercase ASCII letters.

Expansion uses checked arithmetic and rejects missing remaps, interface numbers outside the positive u32 range, uncontrolled interfaces, and duplicate assignments across rules or compute inputs, even when their target states match. A failed expansion returns no partial targets and leaves caller inputs unchanged. Validation and expansion errors identify the rule and field, such as rules[2].terms[0].divisor, without including coordinate values, remap contents, or interface names. Indices are zero-based.

The calculator has no fixed count limits or work budget. Work and memory usage depend on the supplied inventory, controlled interfaces, and calculation rules. Callers are responsible for applying resource policies appropriate to their execution environment. No device state is changed by calculation or validation.