Summary

  • IGMPv1 and v2 spread responses with random timers, then let a host cancel its pending Membership Report when it heard an equivalent report from another member. The router normally learned one fact per group and link: at least one listener existed.
  • IGMPv2 made a Leave message trigger a short series of group-specific queries; it never treated one departure as proof that the group was empty. IGMPv3 later removed host report suppression because source lists, state changes and switch-port forwarding made different hosts' reports no longer interchangeable.

The dangerous moment after a simple question

A multicast router attached to a busy local network has a modest question: for which multicast groups should it continue forwarding traffic onto this link? The obvious procedure is to ask every host and wait for every answer.

It is also a recipe for a burst. If hundreds of machines receive the same Query and answer at once, the control protocol creates the congestion it was meant to avoid. Yet appointing a permanent representative would create a different problem. Group membership is dynamic, attached to interfaces, and not owned by a central member.

RFC 1112, published in August 1989, solved the problem without naming a leader. Its design depended on asking how much evidence the router actually needed. For the forwarding decision then in view, the router did not require a directory of hosts. It needed to know whether at least one member of a group existed on the attached network.

That smaller question made one report sufficient.

Everyone waited, but only the first timer spoke

An IGMPv1 router sent a Host Membership Query to the all-hosts address 224.0.0.1 with an IP time-to-live of one. Each host considered its memberships on the interface where the Query arrived. Instead of answering immediately, it chose a separate random delay between zero and ten seconds for every applicable group and started the timers.

The randomness performed the first compression. Reports were spread across an interval rather than synchronized at the instant of the Query. The destination of each report performed the second. A host sent its report to the multicast address of the group being reported, also with TTL one. Other members of that group on the same link could therefore hear it.

If a host was still waiting when it heard a valid report for its group, it stopped its own timer. In the normal case, the member whose timer expired first sent the only report for that group. The rest did not fail to answer. Their silence was the prescribed acknowledgement that an equivalent witness had already supplied the fact the router needed.

The mechanism has sometimes been described as report suppression, but the word can obscure its precision. It did not suppress membership at the application or host. It suppressed redundant transmissions for one group on one interface during one reporting cycle. Nor did the first reporter become an owner, spokesperson or access authority. It happened to win a randomized race.

Presence was soft state, not a permanent register

The router refreshed its knowledge with periodic Queries. If reports for a group ceased through the specified query and timeout process, it could assume no local member remained and stop forwarding remotely originated traffic for that group onto the link. A crash, lost state or silent departure therefore did not require a durable deregistration transaction. Absence emerged after refreshed evidence failed to arrive.

Joining used the opposite bias. A host that joined a group sent an unsolicited report immediately, then repeated it after short delays to cover possible loss. It did not wait for the next periodic Query, because it might be the first local receiver and traffic might otherwise never reach the link.

This was a deliberately asymmetric system. Presence could be announced promptly by any member. Continued presence could be refreshed by any one member. Absence required time and a failed attempt to elicit a surviving witness.

The economy follows directly from the decision object. If the router stores only “group G is wanted somewhere on this link,” then two hosts' positive reports for G are interchangeable. Suppressing the second report discards no information needed for that decision.

A Leave was evidence to ask, not evidence to stop

RFC 2236 introduced IGMPv2 in November 1997. It added an explicit Leave Group message, a Max Response Time carried in Queries, group-specific Queries, and a common querier-election rule. The point was lower leave latency: stop unwanted forwarding sooner after the final listener departs.

The protocol remained careful about what one host could prove. A departing host should send Leave if it remembered itself as the last host to answer a Query for the group. A host that had not been the last reporter could send nothing, because another member had recently supplied evidence of presence.

Even the last reporter's Leave did not authorize immediate deletion. The querier sent a configured number of Group-Specific Queries at short intervals. If another member remained, it could answer and preserve the group state. Only if no report arrived after the final response interval did the router conclude that the link was empty for that group.

Leave therefore meant “retest this premise now.” It did not mean “trust the departing machine to speak for everyone.” The design shortened the wait while preserving the difference between one departure and collective absence.

Backward compatibility made that boundary visible. When an IGMPv1 member was present, a v2 router ignored Leave messages for the affected group because the older host did not implement the v2 departure procedure. A new fast path could not safely reinterpret an old participant's silence.

Source filtering made one report unlike another

IGMPv3, published in RFC 3376 in 2002 and now specified by RFC 9776, changed the question. A listener could ask for traffic to group G only from listed sources, the INCLUDE mode, or from every source except a list, the EXCLUDE mode. The local system combined socket requests into interface state, and reports could carry that source list.

Router state was no longer merely group G. It became a group plus source-state problem. A v3 report could describe current state, a change of filter mode, newly allowed sources or sources that should be blocked. Two hosts attached to the same link could legitimately want different senders. The first report could not stand for the second without losing the difference.

The v3 design rationale therefore removed host Membership Report suppression. A host no longer cancelled its own v3 report merely because it observed another host's similar transmission. The standard records several reasons: a router might need per-host observations for fast leave or accounting; snooping bridges made suppression troublesome; eliminating the cancellation state simplified hosts; and a single v3 packet could bundle records for several groups, recovering efficiency without erasing distinct state.

The reversal did not abandon pacing. Hosts still distribute responses across the Max Response Time and must not answer a General Query immediately. The problem of simultaneous response remains. What disappeared was the inference that hearing someone else makes one's own evidence redundant.

A switched LAN was no longer one acoustic room

The original suppression algorithm assumed a shared reporting environment: members of the group could overhear the report that caused them to remain silent. Ethernet switching complicated that assumption.

An ordinary bridge floods multicast, but an IGMP-snooping switch inspects membership messages and builds port-specific forwarding state. RFC 4541 documents the resulting tension. A snooping switch commonly forwards Membership Reports toward router ports rather than flooding them to host-only ports. If it did flood a report, a v1 or v2 host on another port might cancel its own report after hearing it. The router had its one group-presence fact, but the switch might never learn that the silent host's port also needed the multicast stream.

The optimization had crossed a boundary. For the router's group-per-link decision, one witness still sufficed. For the switch's group-per-port decision, the witnesses were not interchangeable. A report saved at the IP control layer could become a missing forwarding entry at Layer 2.

IGMPv3 did not make every network device omniscient. It aligned the reporting rule with the finer-grained evidence now being consumed. Reports moved to 224.0.0.22, where v3-capable multicast routers listen, and could carry multiple Group Records. Snooping devices still require correct version handling, timeouts and topology-change behavior.

Compatibility preserved the old meaning of silence

Version transitions do not occur link by link on one clean date. An IGMPv3 system keeps compatibility state after hearing older Queries and can send v1 or v2 reports as required. Routers likewise track older members because their presence changes the safe interpretation of Leave and source-specific state.

This means the wire version is not decorative metadata. One old participant can move a group back to older evidence semantics for a timeout interval. In Source-Specific Multicast, defined by the (S,G) channel model in RFC 4607, losing source-specific evidence can deliver traffic from unwanted sources or prevent the intended pruning. The current v3 standard accordingly says an SSM-aware host must not allow its v3 membership record to be suppressed by an older report.

Version compatibility is useful because it preserves reachability during migration. It also creates an operational cost: the least expressive participant may determine which distinctions the link can safely use.

The report never became permission

IGMP's evidence remains narrow. It signals IPv4 multicast reception interest between systems and neighboring multicast routers. It does not construct the wide-area multicast tree, choose who may transmit, authenticate a subscriber, guarantee packet delivery or establish a durable identity.

RFC 9776 states that IGMP supplies no confidentiality; devices on the link can observe potentially sensitive group interests. Forged reports can keep traffic flowing where no genuine receiver exists, and forged older-version reports can hold a router in a compatibility mode that weakens fast leave or source-specific handling. TTL one, Router Alert rules and on-link source checks constrain some attack paths, but they are not cryptographic proof.

That limit is part of the history, not a defect to hide. IGMP succeeded by standardizing the minimum local evidence routing needed, leaving authorization and application identity elsewhere. Trouble begins when a forwarding hint is promoted into a claim about a person, entitlement or delivery outcome.

Sources and limits

RFC 1112 defines the v1 randomized timers and overheard-report cancellation. RFC 2236 defines v2 Leave processing, last-member queries and older-version coexistence. RFC 3376 records the 2002 v3 rationale; RFC 9776 is the current backward-compatible v3 standard. RFC 4541 supplies informational guidance for snooping switches, and RFC 4607 defines the source-specific service model.

These records establish protocol mechanisms, not adoption rates, product conformance, one universal switch behavior or the identity of any listener. Reading the end of suppression as a change in the value of evidence is an inference from the standards' state models and stated rationale.