Summary

  • IPv4 began with fragmentation inside the network: a gateway could cut an oversized datagram into pieces for a smaller next link, while the destination carried the burden of putting it back together.
  • Path MTU Discovery moved the decision outward. First, routers used ICMP to report the constraining link so the source could shrink; later, packetization layers learned usable size from acknowledged probes even when ICMP disappeared.
  • The enduring design is a division of authority. The path supplies bounded evidence, but the sender keeps the estimate, validates signals, distinguishes probe loss from congestion and changes packet size as routes change.

A packet larger than the next network

The Internet Protocol did not assume that all constituent networks carried the same-sized packet. RFC 791, published in September 1981, called addressing and fragmentation IP's two basic functions. If a datagram left a network that tolerated a large packet and met a smaller network later in its route, an IP module in a gateway could divide it.

The header shows what that promise cost. An Identification value marks fragments as belonging to one original datagram. Fragment Offset places each piece. More Fragments says whether another piece follows. The destination reassembles pieces that share source, destination, protocol and identification. The network can therefore conceal a local packet-size boundary from both application and transport—until a piece goes missing.

Loss does not stay local to the lost fragment. Reassembly needs the set. Transport retransmission may have to send the original data again, producing new fragments. Intermediate fragmentation also means a router performs work and creates more packets at precisely the constricted point. The apparent convenience is distributed state and amplified loss.

IPv4 already included an escape hatch. A sender could set Don't Fragment. If the next link could not carry that datagram intact, the router had to discard it instead of cutting it. But a discarded packet is useful only if the source learns why it disappeared and what smaller size would pass.

The relevant number is not “the Internet's MTU”. Each link has a limit. For one route, the Path MTU is the smallest of those limits. Another route to the same destination may have a different minimum; a tunnel can add headers and narrow the available space; a routing change can invalidate yesterday's result. Packet size is thus path state, not a permanent attribute of a host.

The option every router had to touch

Early discovery proposals tried to make the path write down its own narrowest point. RFC 1063, published in July 1988, proposed two IPv4 options. A Probe MTU option would begin with the first link's limit. Every gateway would compare that value with its inbound and outbound links and lower it when necessary. A Reply MTU option would carry the learned minimum back.

The proposal diagnosed the trade correctly. A fixed small packet wastes capacity and adds headers; a large guess invites fragmentation; extra probes burden the network and become stale as paths change. It also separated two quantities often confused in implementations. TCP Maximum Segment Size says how much TCP payload a peer is prepared to receive. Path MTU describes what the intervening path can carry as an IP packet. One cannot substitute for the other.

Yet the option made discovery depend on every participating gateway recognising and updating a new field. RFC 1063 itself noted that a gateway might not know both interfaces when it processed IP options and that support could require major software changes. The scheme assigned the path excellent knowledge but demanded coordinated participation throughout the path.

Two years later, the standards track chose a smaller change.

Let the narrow link answer back

RFC 1191, published in November 1990, defined classical Path MTU Discovery. A source begins with the MTU of its first hop and sends IPv4 datagrams with DF set. When a router cannot forward one intact, it discards the datagram and sends an ICMP Destination Unreachable message whose code means fragmentation was needed while DF was set. The source then reduces its estimate and sends smaller packets.

The router no longer has to update every packet. It speaks only when a size fails. RFC 1191 used an unused field in that ICMP message to return the next-hop MTU—the limit that caused the discard. A host can converge directly on a useful value, then continue setting DF so a later route to a narrower path produces new evidence.

Deployment could not assume that every router changed at once. An older router returned zero in the new field. For those messages, the source searched downward through plausible MTU “plateaus”: values chosen from common link technologies. The document preferred underestimating by a few percent to overestimating by one octet. It also warned that the table was implementation guidance, not an eternal registry of network sizes.

Discovery has an upward direction too. Routes may improve. RFC 1191 therefore let a host age a cached estimate and occasionally try a larger datagram, with minimum intervals to avoid turning search into repeated loss. An error may reduce the estimate; its contents must never increase it. That asymmetry recognises the evidentiary difference between a claimed limit and successful delivery.

Classical PMTUD altered the location of work. A router still knows the failed outbound link, but the source retains the path estimate and changes future packetization. The common network reports a bounded fact; the endpoint executes the adaptation.

When silence looked like reachability

That division depends on the report reaching the source. RFC 2923, published in 2000, catalogued the resulting “black hole”. Routers sometimes failed to generate the ICMP message because of bugs or configuration. Firewalls often suppressed ICMP indiscriminately. The sender kept transmitting large DF packets, never learning that it needed to reduce them.

The failure looked stranger than an outage. A TCP three-way handshake uses small packets and succeeds. An interactive control connection may work. Pings may return. Then the first bulk data segment crosses the hidden boundary, disappears and is retransmitted at the same unusable size. The connection makes no progress until timeout. The destination is reachable and the path is broken for a particular packet size at the same time.

This was not merely an argument for smaller defaults. Falling back can restore service while masking a misconfigured firewall and imposing a recurring performance penalty. Nor is every missing ICMP message caused by a firewall. Tunnels add headers and redirect errors toward an ingress; asymmetric routing can separate data and feedback paths; ICMP generation may be rate-limited; a link-layer device may drop an oversize frame without creating a usable IP error.

RFC 8201, the current IPv6 PMTUD specification, records the same operational shape: a handshake can complete and data can hang when Packet Too Big messages are blocked. IPv6 makes the responsibility boundary sharper. Routers do not fragment forwarded IPv6 packets. The source chooses the packet size and, if fragmentation is used, creates the fragments.

Proving size from the edge

Classical discovery asks the network to explain a failure. RFC 4821, published in 2007, asks an endpoint to prove success. Packetization Layer Path MTU Discovery begins with a working size and sends selected larger probes. If delivery is confirmed, the lower bound rises. If an isolated probe fails conclusively, the upper bound falls. Repeated tests narrow the interval until further gain is not worth the search.

The packetization layer is the component that decides how application data becomes packets—TCP in the familiar case, but not only TCP. It can know whether a particular probe was acknowledged, keep ordinary traffic at a safe size and treat the experimental loss differently from a general congestion signal.

That distinction must remain narrow. A timeout or losses beyond the isolated probe are inconclusive; normal congestion control still applies. One missing packet cannot announce with certainty that the path is too narrow. It may have met congestion, corruption or a transient route. Likewise, one successful probe proves passage on the observed path at that moment, not permanent capacity across every equal-cost branch.

PLPMTUD can combine ICMP with probes or operate without ICMP. It does not abolish the network's evidence; it removes a single feedback channel as a precondition for progress. The price is more endpoint machinery: per-packet delivery evidence, probe scheduling, search bounds, shared path caches, header accounting and careful interaction with congestion control.

RFC 4821 describes the result as more consistent with the Internet's end-to-end principle. The judgement is less important than the mechanism. Knowledge closest to a link remains useful, but a sender no longer grants an unauthenticated or missing error sole power over its behaviour. It tests what the path actually delivers.

IPv6 and datagram transports complete the direction

IPv6 did not make path size disappear. RFC 8201 defines the same essential minimum across links and relies on ICMPv6 Packet Too Big for classical discovery. A node that declines discovery can stay at the IPv6 minimum link MTU, sacrificing efficiency on paths that support more. A node that sends larger packets without robust feedback can still encounter a black hole.

Datagram transports required a more explicit version of endpoint proof. RFC 8899, published in 2020, defines Datagram PLPMTUD for transports and applications including UDP-based protocols, SCTP and QUIC. Some of those protocols have acknowledgments; bare UDP does not. The application or packetization layer must therefore supply a way to confirm that a probe reached its peer.

The algorithm keeps normal data at or below a current working size, permits specially identified probes above it, lowers the estimate after a detected black hole and searches upward after success. Packet Too Big messages can accelerate the process, but they are optional inputs and must be validated before changing state. The network's message becomes one piece of evidence among delivery observations.

Across the sequence, responsibility did not jump once from core to edge. It moved by correction. Router fragmentation hid heterogeneity but multiplied work. An in-packet option asked every gateway to cooperate. ICMP-based discovery reduced router involvement but assumed a reliable return signal. Packetization-layer discovery used endpoint-controlled experiments to survive that signal's absence. Each step retained something from the previous design and narrowed what had to be trusted.

The packet as a bounded decision

Path MTU is often taught as a number—frequently 1500, sometimes 1280, occasionally a tunnel-adjusted value. Its history is better understood as a control problem. Which component can observe the constraint? Which evidence can cross administrative boundaries? Which actor pays when the evidence is wrong? Who may revise the decision?

A link operator controls a local MTU. A router enforces it and may report it. A firewall can accidentally erase the report. The source controls packetization, caches a path estimate and observes delivery. The application may have to make a datagram probe recognisable. No entity owns the whole path, and no observation remains true indefinitely.

The stable common rule is therefore modest: do not make an intermediate network silently carry unlimited adaptation; expose a bounded failure when possible; let the endpoint test, remember and revise the size it uses. The packet did not acquire intelligence. The architecture learned to keep the decision where evidence could be combined with consequences.

Sources and evidence limits

The original IPv4 fragmentation model and DF behaviour come from RFC 791. The gateway-updated option proposal is RFC 1063. Classical PMTUD, next-hop MTU reporting, old-router plateaus and cache aging are specified in RFC 1191. The TCP black-hole account and its diagnostic limits come from RFC 2923.

RFC 4821 supplies packetization-layer probing and its congestion-control boundary; RFC 8201 supplies the current IPv6 PMTUD model; RFC 8899 extends robust probing to datagram transports. These documents establish specifications, design debates and recognised failure modes. They do not establish a single inventor, a universal adoption date, the prevalence of each failure or the current configuration of every network.