RFC 3116: The Clock Inside an ATM Delay Result

Summary

RFC 3116 required correlated transmit and receive timestamps driven by a common clock signal, with timestamp width, accuracy and rollover range disclosed; a delay number without that instrument record was not comparable evidence.
The same switch could produce different distributions under one, twelve or many virtual circuits, steady or bursty load and different bearer classes. The graph described a controlled trial, not an intrinsic promise about production service.

The subtraction concealed an apparatus

Cell transfer delay sounds like a subtraction: note when one ATM cell leaves a first measurement point, note when its corresponding arrival is observed at a second, then take the difference. RFC 2761 supplied that definition. RFC 3116, published as an Informational methodology in June 2001, supplied the apparatus that made the subtraction defensible.

The transmitter and receiver needed time information that could be correlated. RFC 3116 recommended feeding both from one common clock signal, at least 100 MHz, expressed in the document as 10-nanosecond resolution. The preferred test unit was an O.191 cell carried inside a valid IP packet. If another test cell was used, its construction could not remain invisible.

This was not merely a demand for more decimal places. Two counters can each be precise and still disagree. One can also wrap from its maximum value back to zero while the other event is recorded on the opposite side of that boundary. A naive subtraction then converts a later arrival into an impossible negative interval or an enormous delay. RFC 3116 therefore required the maximum timestamp value to be recorded and required the test-cell description to state timestamp length, rollover value and accuracy in nanoseconds.

The performance claim thus began before the cell entered the switch. It began with the clock source, timestamp capture points, counter width and rule for matching a transmitted cell to its received observation. The switch under test did not own the stopwatch.

One clock did not erase every uncertainty

A common frequency reference removes an important class of disagreement, but it does not magically prove the whole timing chain. The clock may feed two counters, while their capture circuits observe different physical points. Timestamp insertion can precede the actual exit event; receive capture can follow the actual entry event. Counter phase, wiring, calibration and correlation logic remain part of the instrument.

The contemporaneous IP Performance Metrics work made this boundary even more explicit. RFC 2679 separated synchronization, accuracy, resolution and skew. It explained that synchronization error and the resolution of both clocks enter the uncertainty of a one-way-delay observation, while the gap between host time and wire time contributes another error. RFC 2330 supplied the broader measurement framework. RFC 7679 later replaced RFC 2679 without turning a timestamp into self-authenticating truth.

Those IPPM documents measured Internet packets rather than ATM cells in a device test. They should not be collapsed into RFC 3116. Their value here is historical corroboration: at the same moment, two IETF measurement traditions were refusing to let the instrument disappear behind the metric.

RFC 3116's own rule was narrower and concrete. Describe the cell. Preserve enough counter information to resolve rollover. State the time accuracy. Correlate the transmit and receive observations. Only then calculate the interval.

Traffic was part of the result

Even a trustworthy clock does not make delay independent of load. ATM switches multiplexed cells from virtual connections, scheduled bearer classes and buffered bursts. RFC 2761 defined Cell Delay Variation as the spread of cell-transfer delay under a specified traffic load, orientation, distribution and integration period. Change those conditions and the distribution changes.

RFC 3116 accordingly repeated its CTD procedures across one VCC, twelve VCCs and the maximum number the system supported. It distinguished constant, bursty UBR, VBR and mixed loads. It required packet size, packet rate, bearer class, VPI/VCI and the PCR, SCR and MBS traffic descriptors to accompany the result. For tests that were not throughput tests, the offered rate should not exceed 90 percent of line rate.

Before accepting the timing run, the tester compared packet counts to verify connectivity and load. If the system did not transmit the expected count, the offered rate was reduced. This step did not prove that every internal queue behaved correctly. It prevented a delay experiment from quietly becoming a loss or overload experiment while retaining the label “delay.”

The reporting format mattered for the same reason. Text gave the configuration; a graph exposed time behavior; a histogram exposed the distribution. An average alone could hide a long tail, while a maximum alone could hide how rarely it occurred. RFC 3393 and RFC 5481 later sharpened distinctions among packet-delay-variation formulations. A shared unit did not make two aggregations interchangeable.

The trial had a before and an after

RFC 3116 also controlled the temporal boundaries around measurement. A trial could first send a PNNI routing update and wait for routing to settle. It then performed ATMARP resolution under the Classical IP over ATM rules of RFC 2225, applied the load, waited for residual packets, and allowed the system to restabilize before the next trial.

These phases prevented state preparation from being mistaken for steady forwarding. Address resolution proved that an IP destination could be mapped to an ATM address for the test. It did not prove the later delay distribution. A routing wait reduced one source of transient state. It did not certify that every control-plane transition had completed. A residual-packet wait kept late observations from leaking into the next trial. It did not transform late traffic into a successful result.

The baseline stimulation period was at least 60 seconds. When variance remained high, the methodology recommended at least 300 seconds. The longer run was not a ritual guarantee of truth; it was an admission that a small observation window could overstate stability. Repeatability, variance and statistical significance still required judgment.

RFC 1242 and RFC 2544 had already established BMWG's division of labour: terminology defined what a benchmark meant, while methodology controlled how it was collected and reported. RFC 3116 applied that discipline to an unusually layered system. Physical SONET behavior, ATM cells, AAL5 packets, signalling and ILMI could each offer a plausible number. The test had to say which layer, which event and which conditions the number represented.

An aggregate could hide the culprit

The methodology could extend beyond one device and treat several interconnected systems as a single system under test. That made end-to-end empirical characterization possible across heterogeneous apparatus. It also weakened attribution.

RFC 3116 warned that an aggregate benchmark might conceal asymmetries between devices or latency introduced by other equipment. Two systems could produce the same total while distributing delay differently. They could also produce different totals because their supposedly comparable configurations lacked a common denominator.

This is the difference between measurement and blame. A correlated timestamp pair shows elapsed time across the declared boundary. It does not, by itself, locate a queue, identify a scheduling decision or prove which component should be repaired. Internal attribution requires more observation points or a controlled intervention. Production attribution requires still more: current configuration, path, load and operating state.

The RFC contained no named vendor ranking and no deployment survey. It defined how vendors and users could produce more comparable laboratory data. Turning that method into a claim that one historical switch was faster than another would add evidence the document does not contain.

The number was a chain of custody

The durable lesson is not limited to ATM. A benchmark value is a compressed record of actors and transitions. The generator chose a load. The clock supplied a reference. Capture hardware assigned event times. Correlation logic paired observations. The counter encoding preserved—or failed to preserve—an epoch. The switch forwarded under a particular VCC and bearer configuration. The reporter reconstructed, grouped and plotted the samples. The reader decided whether another trial had sufficiently similar denominators.

If any link disappears, the number can remain syntactically valid while losing its authority. A ten-nanosecond tick is not ten-nanosecond end-to-end accuracy. A common clock is not proof that event capture occurred at the intended wire points. A histogram is not a service-level agreement. A clean laboratory interval is not application experience. A published methodology is not evidence that a vendor followed it.

This is the measurement version of a wider governance principle: participation is not authorization, and representation is not proof of the represented reality. The timer participates in the claim but cannot authorize the conclusion by itself. The result becomes usable when the boundary, instrument, transformation and context remain attributable.

RFC 3116 made an ATM delay graph harder to publish casually. That was its achievement. It insisted that the apparent fact on the y-axis carry the clock, counter and traffic history that produced it.

A Delay Number Was Also a Clock Claim: RFC 3116

Why it matters

Summary

The subtraction concealed an apparatus

One clock did not erase every uncertainty

Traffic was part of the result

The trial had a before and an after

An aggregate could hide the culprit

The number was a chain of custody

Sources

Deeper Profile Context

Strategic Circle

Leadership Alliance

Strategic Circle Briefing

Leadership Alliance Briefing