Summary

  • RFC 3497 extended RTP sequencing to 32 bits because a 16-bit counter would wrap in 336 milliseconds while carrying 1.485 Gbit/s of uncompressed HDTV.
  • The extension could reveal loss and disorder; it could not reduce the fixed sending rate. On unsafe best effort, the receiver's only compliant congestion response was to leave.

At one thousand bytes per packet, the familiar 16-bit RTP sequence number lasted 336 milliseconds. That was too short to tell a late packet from a counter that had already circled back. RFC 3497 added another sixteen bits in a four-octet payload header. The resulting 32-bit sequence space lasted roughly six hours.

It was a beautiful measurement fix. It did not regulate one bit.

The source was SMPTE 292M, the serial interface used to move uncompressed high-definition television among cameras, encoders and recorders. It carried ten-bit words at 1.485 Gbit/s, or the closely related 1.485/1.001 rate. A dedicated cable made sense inside a studio. The RFC asked how the same signal could cross a wide-area IP network without losing the boundaries that made it television.

Each video line could occupy several RTP packets. Start-of-active-video and end-of-active-video timing structures could not be split, because the decoder used them to find scan lines. The ordinary RTP header was followed by the payload header, and the marker bit identified the packet ending a frame. An eleven-bit line number gave the application another way to recover its place if the packet carrying the normal end marker disappeared.

The format also respected pixel groups. A common 4:2:2 representation aligned on five-octet groups; 4:2:0 and 4:4:4 used fifteen. A pgroup of one permitted arbitrary octet boundaries when the source format was unknown, but the crucial SAV and EAV structures still remained indivisible. video/SMPTE292M, a required clock rate and an optional pgroup gave SDP enough information to describe the mapping.

That solved representation. The network problem was of another kind.

The RTP clock ran at 148.5 MHz, allowing a receiver to reconstruct sample timing without first knowing the precise HDTV source format. Its 32-bit timestamp wrapped in 21 seconds. RTCP's sender octet count wrapped in 23 seconds, and its cumulative-loss field in 93. Applications wanting lifetime totals had to count those rollovers themselves.

One stream therefore contained several notions of continuity: six hours for extended packet sequence, 21 seconds for media time, 23 seconds for reported octets and 93 seconds for reported loss. Confusing any wrap with a reset would corrupt the evidence while the packets remained perfectly conformant. RFC 3497 demanded stateful observation precisely because the line rate made bounded counters turn quickly.

Observation was not capacity. The RFC said the constant-rate stream was not congestion-controlled and warned that 1.485 Gbit/s before RTP overhead could deny service on most Internet paths then available. Safe use had to be tightly constrained to private networks, suitably connected endpoints or networks where reservation and enhanced quality of service were real.

Even a QoS label was not enough. A receiver using enhanced service should monitor loss to verify that the requested service was actually being delivered. If observation contradicted the promise, it should assume best effort. The symbolic claim changed nothing about the execution layer; packet loss was the stronger receipt.

Best effort produced the hardest rule. The receiver had to monitor loss and leave the session when it became too high. The RFC defined acceptable loss by asking whether a TCP flow on the same path, under the same conditions, could achieve average throughput at least as high as the RTP flow. The HDTV sender could not gradually back off to satisfy that comparison. Once the path failed the test, receiver departure was the only available circuit breaker.

That is why the 32-bit sequence number matters historically. It made an enormous, inelastic flow auditable. It could show exactly which packets vanished, while refusing to pretend that a more accurate ledger manufactured bandwidth. RTP carried media; RTCP reported conditions; SDP described the session; IANA coordinated names. None of those authorities installed a queue or widened a link.

Later work sharpened the surrounding architecture. RFC 3550 replaced the original RTP specification. RFC 4175 generalized uncompressed-video carriage. RFC 8083 formalized RTP circuit breakers, RFC 8085 gave broader UDP guidance, and RFC 8888 added richer congestion-control feedback. Those tools improve evidence and response. They do not retroactively make RFC 3497's fixed-rate payload elastic.

The receipt chain remains strict: the session is described; packets are formatted; scan-line boundaries survive; gaps are observable; clocks are reconstructed; RTCP totals are made rollover-aware; promised QoS is measured; TCP-equivalent coexistence is tested; useful video arrives; and the session remains safe. A packet trace can prove the first five while the shared path fails the last five.

RFC 3497's discipline lay in saying so. It offered a minimum interoperable mapping, not a mandate for deployment. It separated the symbolic precision of counters from the physical truth of capacity. When the two disagreed, the receiver was instructed to believe the path and leave.

Sources