What determines how fast a routed network converges after a failure, and how do you make it faster?
Show the answer
Say this first
Convergence is detection, plus propagation, plus computation, plus updating the FIB. Detection is usually the biggest piece: a lost carrier is noticed fast, but a silent failure waits for hello timers (OSPF 40 s dead time on LAN, EIGRP 15 s hold, BGP 180 s hold), which is why BFD exists. After that come flooding and SPF or DUAL timers, and the time to rewrite thousands of FIB entries. Precomputed backups (EIGRP feasible successors, LFA and TI-LFA, BGP PIC) remove most of the computation and FIB time.
Go deeper
The four parts
| Part | What drives it | How to shrink it |
|---|---|---|
| Detection | carrier loss, protocol hello and dead timers | BFD (for example 50 ms x 3 = 150 ms), carrier-delay tuning, point-to-point links |
| Propagation | flooding and update pacing | LSA and SPF throttle timers, fewer hops, summarization |
| Computation | SPF, DUAL, BGP best path | feasible successors, LFA/TI-LFA, incremental SPF |
| FIB update | number of prefixes to rewrite | hierarchical FIB and BGP PIC: change one shared next hop entry |
Typical config pieces on IOS-XE
interface GigabitEthernet0/0/1
bfd interval 50 min_rx 50 multiplier 3
ip ospf network point-to-point
!
router ospf 1
bfd all-interfaces
timers throttle spf 50 200 5000timers throttle spf 50 200 5000 runs the first SPF 50 ms after a change, then backs off up to
5 seconds if changes keep coming.
Why point-to-point helps
On an Ethernet link between two routers, OSPF network type point-to-point skips DR election and the wait timer, so adjacencies form and recover faster.
Precomputed backups
- EIGRP feasible successors are installed with no queries.
- OSPF and IS-IS loop-free alternates (and TI-LFA with segment routing) preinstall a backup next hop.
- BGP PIC Edge (
bgp additional-paths install) preinstalls a backup path, and the hierarchical FIB switches every prefix that shares a next hop in one step.
The side effect to mention
During link state convergence, routers finish at slightly different times and can loop packets briefly between them (microloops). Fast reroute and ordered FIB updates reduce them.
They will ask next
Why not just set the OSPF hello to 1 second everywhere?
Fast hellos cost CPU on every interface and are processed by the control plane, so a busy CPU can miss them and drop adjacencies. BFD is lightweight, often runs in hardware or a dedicated process, and is shared by all protocols on the link.
Traps
- Talking only about SPF speed. On most networks detection dominates.
- Tuning timers aggressively without dampening. Fast reaction to a flapping link can make the whole network unstable.