What it is
In a gossip protocol every node periodically talks to a few random peers, and news spreads like an epidemic: the number of informed nodes roughly doubles each round, so a rumor reaches n nodes in about log2 n + ln n rounds with push gossip. SWIM applies the same idea to membership: each period a node pings one member, asks a few others to ping for it when there is no answer, marks the target suspect if that also fails, and declares it dead only if the target has not refuted the suspicion with a higher incarnation number in time. Updates ride piggybacked on the probe messages, so per-node traffic stays constant as the cluster grows.
When to use it
Use it to see why gossip finishes in logarithmic time and how push, pull and push-pull differ at the start and at the end of a spread; how quickly a crash is noticed and how long agreement takes; why a few lost packets cause false suspicions unless indirect probes are on; how a longer suspicion timeout trades slower death verdicts for fewer false ones; and what a partition does to each side’s view and how refutations repair it after the heal.
Gotchas & limits
This is a teaching model, not an implementation of memberlist or Serf. Joins, anti-entropy over TCP, Lifeguard and message size limits are not modelled, and a periodic gossip to nodes believed dead stands in for state synchronisation after a partition. Periods and timeouts are illustrative, and the retransmit and suspicion multipliers are recalled from the SWIM paper and common implementations, not re-checked. Results show the mechanism and its scaling, not a prediction for your cluster.