Summary

  • RFC 9571 extends RFC 6374 measurement to generic MPLS and multipoint-to-point LSPs by using Synonymous Flow Labels to identify packet batches.
  • A label assigns packets to a cohort; accurate loss still depends on draining reordering, interfaces, processor cores and forwarding pipelines before the Query samples the receiver.
  • A defensible result needs separate receipts for scope, batch membership, counter closure, overlap reconciliation, method and configuration epoch, returned statistics and later service outcome.

The color changed before the counter settled

Imagine a busy MPLS edge alternating two labels across one production flow. The first label marks one loss interval, the second marks the next. Sender counters look clean. Receiver counters are exposed through an OAM response. The subtraction says three packets vanished.

But one packet from the first interval was still crossing a forwarding pipeline when the receiver read its counter. Another arrived out of order through an equal-cost path. A third reached a different interface whose accounting state had not yet synchronized. Nothing in the label itself distinguishes actual loss from an observation taken before the window closed.

RFC 9571 is valuable because it makes that distinction explicit. It extends RFC 6374 performance measurement from MPLS Transport Profile into generic MPLS and multipoint-to-point cases. Synonymous Flow Labels give packets stable batch membership. The sender still waits until it is confident that every packet in the batch could have arrived before asking for the receiving count. Identity solves one problem. Completion remains an operational judgment.

The old delimiter was a packet in the same uncertain stream

RFC 6374 can use a Loss Measurement message as the accounting boundary. RFC 9571 lists why that boundary can drift. ECMP can deliver data out of order with respect to the LM message. Packets may land on different interfaces. Work may be distributed across processor cores. Link aggregation can separate the relevant events. A forwarder may have a long distance between handling a packet and incrementing its counter.

These are not edge cases around an otherwise exact statistic. Modern networks often lose very few packets, so one packet assigned to the wrong interval can dominate the result. A delimiter that arrives is not proof that all earlier data-plane work has become visible in the counter being read.

Alternate marking changes the evidence model. Rather than ask one packet to divide time everywhere, the sender marks a population so adjacent batches can be recognized. RFC 9571 carries that model through SFLs. The label is a membership receipt: this packet belongs to batch A. It is not a settlement receipt: batch A is now complete at every observation point.

Multipoint measurement needs a source, not merely a total

Generic MPLS introduces another scope problem. In a multipoint-to-point LSP, the MPLS packet does not provide the ordinary source address that an operator would use to separate ingress populations. A sink can hold a perfectly accurate aggregate and still be unable to answer how many packets were lost between one source and that destination.

RFC 9571 therefore needs source-specific information alongside the batch distinction. This is a recurring operational lesson: the number is only as useful as the population it describes. “Packets received on this LSP” is not interchangeable with “packets received from ingress X during window A.”

The evidence record should bind the FEC, source, destination, SFL allocation, batch index, measurement session and counter location. If any of those coordinates are recovered later from convention instead of retained with the result, the statistic becomes harder to audit precisely when an anomaly matters.

One packet can belong to two statistical stories

Loss windows tend to be long. Delay windows may be short and frequent. RFC 9571 allows the measurements to run concurrently, but calls them only relatively independent. Their batch times can overlap.

A delay-marked packet can also belong to a loss batch. The implementation may update two counters or reconcile them during reporting; that choice is local. If a delay batch crosses the boundary between loss batches A and B, the operator must decide which loss population receives those packets. The Query must then wait until all packets relevant to that decision have reached the receiving LSR.

RFC 9571 constrains the default case deliberately. A delay-capable system must also be able and configured to measure loss. A delay interval must finish inside its enclosing loss batch. A second delay batch must last long enough for the first batch's packets to arrive before the second completes. These rules are not protocol decoration. They define which counter combinations can be reconciled without inventing history.

A zero can mean “not ready,” not “nothing happened”

The document defines several measurement products. Time buckets count packets within configured inter-packet intervals. When those bucket boundaries are first created or changed, the responder must return zero counts until it has been configured for a full measurement period. That zero is a state marker. Reading it as observed zero jitter would turn incomplete configuration into a performance claim.

The classic-statistics form returns packet count, sum of delays, minimum, maximum and a sum of squared inter-packet delays. The analytics system can derive a mean and variance, but only for that batch and formula. Average-delay reporting may instead return summed timestamps or first and last timestamps with a count. Sampled measurement observes selected packets and defines its interval between readout packets, not simply by how long one color was present.

These outputs cannot be flattened into one generic “latency” field without losing method and scope. The receipt needs the ACH type, requested method, response form, batch, timestamp format, configuration epoch and analytics transformation. RFC 9571 also states that combined loss/delay measurement using SFL is not currently supported. A field layout that appears capable of carrying something is not permission to claim the mode exists.

Context sometimes has to be carried because hardware discarded it

The Query follows the data-service MPLS stack and adds the Generic Associated Channel Label so the receiver can distinguish OAM from normal packets. The ACH type names the measurement. But some hardware removes the MPLS stack before handing the remainder to the OAM processor.

RFC 9571's SFL TLV can restore the missing association. It carries the SFL, the FEC used to allocate it, a management batch identifier and an index into the SFLs assigned for that FEC. The index is a convenience agreed during configuration, not a universal identity. The TLV is required over UDP and otherwise must be included unless some method outside the RFC establishes that the responder does not need it.

Return information has its own boundary. In-band point-to-point operation may know where the response goes. Other cases need explicit return context. A valid measurement body without a routable response path is unfinished work, just as a delivered response without the right SFL/FEC association can be a precise answer to the wrong question.

Measurement truth stops before service truth

Suppose every protocol step works. The batch is scoped, the pipelines are drained, the response is authentic, and the analytics formula is correct. The result proves a loss or delay property for that selected population, at those observation points, during that interval, under that method.

It does not prove why a packet was lost, whether an application retransmitted, whether another path performed differently, whether an SLA definition used the same population, or what a customer perceived. Those require additional evidence. RFC 9571 provides stronger network measurement machinery; it does not collapse measurement into causation or outcome.

Privacy and integrity also stay visible. Source or flow granularity can expose more identity. Control traffic can be encrypted and synonymous labels rotated or used concurrently. A control protocol that requests SFLs must establish the request's legitimacy. Corrupting the SFL TLV can disrupt a measurement without the RFC 6374 responder detecting the alteration. An apparently coherent response is therefore not self-authenticating provenance.

Sources

Sources