Factors affecting convergence
In one line
The things that slow or break convergence beyond raw timers - so a design that looks fast on paper can still lose traffic.
The factors
- Recursion - a route whose next hop must itself be resolved (e.g. BGP next-hop via the IGP) adds a dependency; next-hop tracking and BGP PIC reduce its impact.
- Microloops - transient loops while nodes update the FIB asynchronously during reconvergence; mitigated by TI-LFA / ordered FIB.
- Microbursts - sub-second buffer overruns from bursty traffic that drop packets even when average utilization is low; a buffer/QoS problem, not a routing one.
- Physical failures - silent or unidirectional failures that link-state alone misses; need BFD/UDLD and sensible carrier-delay/debounce.
Machine-level overload and positive feedback loops
Convergence has a precise meaning: treat the control plane as a distributed database - the network is converged when every device's forwarding view is consistent and loop-free, not when every database is identical. Three things overwhelm that process at machine level: processing and memory load (bigger tables cost CPU and RAM), the rate of state change (every change forces recomputation, and if changes arrive faster than the control plane can compute and propagate, the network never converges - the global Internet's BGP lives in this condition permanently), and positive feedback loops - a failure raises churn, churn raises CPU load, loaded routers miss hellos, and missed hellos create more failures. Feedback loops are why the remedies are structural rather than faster CPUs: summarization caps how far churn travels (Route aggregation and summarization), flooding-domain boundaries and hierarchy cap who must recompute (Network hierarchy and topologies), and dampening / exponential backoff breaks the loop by refusing to react at the churn's own frequency.
Why it matters for design
You must design for these explicitly. "Sub-second convergence" is meaningless if a unidirectional fault goes undetected, a microloop drops packets during the transition, or a microburst overruns buffers.
Related
Spaced repetition
A transient loop formed while routers update their FIBs asynchronously during reconvergence is a [...] (mitigated by TI-LFA / ordered FIB).
A transient loop formed while routers update their FIBs asynchronously during reconvergence is a microloop (mitigated by TI-LFA / ordered FIB).
Microbursts cause loss even at low average utilization; the fix is a [...] design, not a routing change.
Microbursts cause loss even at low average utilization; the fix is a buffer / QoS design, not a routing change.
Detecting a silent or unidirectional link failure that link-state misses needs [...].
Detecting a silent or unidirectional link failure that link-state misses needs BFD / UDLD.
What is a positive feedback loop in a control plane, and what breaks it?
Failure -> churn -> CPU load -> missed hellos -> more failures. It is broken structurally: summarization caps how far churn travels, flooding-domain boundaries cap who recomputes, and dampening/backoff stops the control plane reacting at the churn's own frequency.
Sources
- Cisco Press, Optimal Routing Design.