EIGRP metric in depth: K values, the formula, wide metrics and unequal-cost load balancing

The K values

The EIGRP metric can use four properties of a path: bandwidth, delay, load and reliability. Five weights, called K values, decide how much each counts. The defaults are K1 (bandwidth) = 1, K2 (load) = 0, K3 (delay) = 1, K4 = 0 and K5 = 0, which means only bandwidth and delay are used. The full classic formula is:

metric = 256 x ( K1 x BW + (K2 x BW) / (256 - load) + K3 x delay )
         ... then multiplied by K5 / (reliability + K4) only if K5 is not 0

With the defaults this simplifies to 256 x (BW + delay), where BW is 10,000,000 divided by the slowest bandwidth on the path in kbit/s, and delay is the sum of the outgoing interface delays in tens of microseconds. Load and reliability change from moment to moment, so using them would make routes change constantly. That is why they are off by default, and why changing K values is almost never a good idea. K values must match between neighbours; you change them with metric weights 0 K1 K2 K3 K4 K5.

A worked example

A route crosses two Fast Ethernet links (100,000 kbit/s, delay 100 microseconds each) before reaching the destination.

1. Bandwidth term: 10,000,000 / 100,000 = 100.
2. Delay term: 100 + 100 = 200 microseconds, which is 20 tens of microseconds.
3. Add them: 100 + 20 = 120.
4. Multiply by 256: 30,720.

Because the metric depends on the bandwidth and delay configured on interfaces, you can steer EIGRP by changing them. Change delay rather than bandwidth when you only want to influence routing, because bandwidth is also used by QoS and by EIGRP's own pacing.

Wide metrics

The classic formula stops distinguishing links above 1 Gbit/s well, because the bandwidth term becomes tiny and interface delays are similar. Named mode EIGRP uses wide metrics: 64-bit values calculated from throughput and latency in picoseconds, so 10, 40 and 100 Gbit/s links get different metrics. The large value is scaled down with metric rib-scale (default 128) to fit the routing table. Routers using wide metrics work alongside classic routers, but for consistent results move the whole domain to named mode.

Unequal-cost load balancing with variance

Like other protocols, EIGRP shares traffic over equal-cost paths, up to four by default (maximum-paths). Its special feature is the variance multiplier. With variance 2, EIGRP also installs any path whose metric is less than twice the best metric, as long as it is a feasible successor. That condition guarantees the extra path cannot loop.

router eigrp 100
 variance 2
 traffic-share balanced
 maximum-paths 4

With traffic-share balanced (the default), traffic is spread in inverse proportion to the metrics, so the better path carries more. Check the shares with show ip route 10.9.0.0, which shows a traffic share count for each path. A neighbour that is not a feasible successor is never used, whatever the variance.

A related setting, ip bandwidth-percent eigrp, limits how much of an interface's bandwidth EIGRP's own packets may use (50 per cent by default). Raise it only on links where the configured bandwidth is lower than the real speed.

Real-world example: two unequal links to a clinic

A healthcare provider connected a clinic with a 100 Mbit/s fibre and a 50 Mbit/s radio link, both terminating on the same pair of Cisco routers. Only the fibre carried traffic, and staff complained about slow image downloads at busy times. The radio path had a feasible distance below twice the fibre path's, and its reported distance was below the fibre's feasible distance, so it qualified as a feasible successor. Setting variance 2 on both routers put the radio link into use, carrying roughly a third of the traffic, and it still acted as an immediate backup if the fibre failed.