Cost and ECMP

Each OSPF interface has a cost, and a route's metric is the sum of the outgoing interface costs along the path. Lower is better. Costs are per direction: R1's cost towards R2 can differ from R2's cost towards R1, which can make traffic paths asymmetric.

By default the cost is derived from bandwidth:

cost = reference bandwidth ÷ interface bandwidth (minimum 1)

The default reference bandwidth is 100 Mbps — a value from the 1990s. With it:

Interface	Cost with 100 Mbps reference	Cost with 100 Gbps reference
10 Mbps	10	10000
100 Mbps	1	1000
1 Gbps	1	100
10 Gbps	1	10
100 Gbps	1	1

Everything from 100 Mbps upward costs 1, so OSPF can't tell a 1G link from a 100G link and picks paths by hop count. Fix it with auto-cost reference-bandwidth 100000 (in Mbps) — on every router, or costs become inconsistent — or set ip ospf cost explicitly on each interface, which many operators prefer because the design is then visible in the config.

Equal-cost multipath

When several paths have exactly the same total cost, OSPF installs all of them (up to a platform limit, maximum-paths in FRR). The forwarding plane hashes each flow onto one of the next hops, so a single TCP connection stays on one path while load spreads across many flows. Data-centre designs depend on this.