Network Working Group S. Dikshit
Internet-Draft HPE
Intended status: Informational 2 November 2026
Expires: 6 May 2027
Operational Telemetry Comparability for IP Performance
Measurements
draft-dikshit-bpm-telemetry-comparability-00
Abstract
Operational telemetry is often used to compare network performance
across observation points, traffic classes, service instances, paths,
devices, and time intervals. The numeric value of a metric is not
sufficient for such comparison unless the observation context and the
comparability conditions are also known.
This document describes a common interoperability problem in
operational measurement systems: values such as delay, loss, ECN
marking, queue occupancy, packet counts, byte counts, and flow
correlation identifiers can be exported from different observation
points or under different aggregation rules while still appearing to
be directly comparable. A collector or controller can therefore
combine values that do not describe the same measurement domain.
The document defines requirements for describing the observation
point, metric source, aggregation interval, traffic class, flow
scope, sampling rule, timestamp basis, freshness, and comparability
domain associated with operational telemetry used for IP performance
measurement. The goal is to make operational measurement results
easier to compare without prescribing one measurement protocol, one
data model, or one congestion control mechanism.
Status of This Memo
This Internet-Draft is submitted in full conformance with the
provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering
Task Force (IETF). Note that other groups may also distribute
working documents as Internet-Drafts. The list of current Internet-
Drafts is at https://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months
and may be updated, replaced, or obsoleted by other documents at any
time. It is inappropriate to use Internet-Drafts as reference
material or to cite them other than as "work in progress."
This Internet-Draft will expire on 6 May 2027.
Copyright Notice
Copyright (c) 2026 IETF Trust and the persons identified as the
document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal
Provisions Relating to IETF Documents (https://trustee.ietf.org/
license-info) in effect on the date of publication of this document.
Please review these documents carefully, as they describe your rights
and restrictions with respect to this document. Code Components
extracted from this document must include Revised BSD License text as
described in Section 4.e of the Trust Legal Provisions and are
provided without warranty as described in the Revised BSD License.
Dikshit Expires 6 May 2027 [Page 1]
Internet-Draft Operational Telemetry Comparability November 2026
Table of Contents
1. Introduction
2. Problem Statement
3. Terminology
4. Motivating Cases
4.1. ECN and Congestion Telemetry
4.2. Queue and Delay Measurements
4.3. AI Network Fabric Measurements
4.4. Flow Correlation Across Observation Points
4.5. Metric Freshness for Traffic Steering
5. Comparability Requirements
6. Relationship to Existing Work
7. Security Considerations
8. IANA Considerations
9. Acknowledgements
10. References
1. Introduction
IP performance measurements are increasingly consumed by automated
systems. Collectors, controllers, service assurance systems, and
traffic steering systems compare telemetry from multiple devices,
paths, and service locations. Such systems often receive numeric
values that appear to be comparable: delay values, packet loss
counters, ECN marking counters, queue depth measurements, packet and
byte counts, or flow identifiers.
In practice, these values are not self describing. A delay value may
be measured at an application boundary, an RDMA completion boundary,
a packet forwarding boundary, or a device internal observation point.
An ECN marking count may be collected at ingress, egress, or an
intermediate point. A queue measurement may refer to a port, a
traffic class, a virtual output queue, or an implementation specific
internal queue. A loss counter may include or exclude retransmitted
traffic. A flow identifier may be unique in one observation domain
but reused in another.
These differences matter because a consumer can reach a wrong
operational conclusion while every exported value is syntactically
valid. The problem is not the encoding of the value. The problem is
the missing statement of the measurement domain in which the value is
meaningful and comparable.
This document defines requirements for operational telemetry used as
input to IP performance measurement. It does not define a new metric
and does not replace protocol specific data models. Instead, it
describes the context that needs to accompany measurements when those
measurements are compared, aggregated, or used by automated systems.
Existing IPPM work already requires metrics to be concrete, well
defined, repeatable, and useful to operators and users. RFC 7799
distinguishes active, passive, and hybrid methods and points to the
importance of observation points or measurement points. IPFIX and
PSAMP define observation points, observation domains, packet
selection, and scoped fields. IOAM defines data fields collected
inside a limited domain and uses namespace context to interpret
fields such as node identifiers, timestamps, and queue information.
Alternate Marking shows why counters from different measurement
points need aligned packet blocks or marking periods before they can
be compared.
This document builds on those ideas. It provides a checklist for a
consumer that asks a simple operational question: given two telemetry
values, what context is needed before the values can be compared,
joined, aggregated, or used as input to an automated decision?
2. Problem Statement
Two telemetry values can have the same metric name and the same unit
while still describing different measurement domains. For example,
two delay values may both be expressed in microseconds, but one may
measure endpoint processing and fabric forwarding while the other
measures only device boundary delay. Two ECN counters may both count
marked packets, but one may represent ingress observations while the
other represents egress observations after remarking, tunneling, or
encapsulation changes.
A collector that receives these values needs to know whether they can
be compared. A controller that uses these values to make a steering
decision needs to know whether a stale service metric and a fresh
network metric can be combined. A troubleshooting system needs to
know whether a counter reset, missing record, traffic class change,
or observation point change created a new measurement series.
Current telemetry work often defines the metric or the encoding, but
the comparability rule is left implicit or is distributed across
protocol specific context. This leads to at least five operational
risks:
1. Incorrect aggregation: values from different observation points
are summed or averaged as if they described one measurement
domain.
2. Incorrect comparison: values collected at different boundaries or
under different traffic class rules are ranked as if they were
equivalent.
3. Incorrect correlation: flow or measurement identifiers are joined
across domains where they are not unique or do not identify the
same observation.
4. Incorrect automation: controllers react to measurements that were
collected at different times or with different aggregation rules.
5. Incorrect fault isolation: endpoint delay, queueing delay, and
fabric forwarding delay are mixed without a way to identify which
component produced the change.
A specification can avoid these risks by making the measurement
domain explicit. The same metric name and unit are not enough. The
consumer also needs to know where the measurement was made, which
packets or flows were included, which time interval was used, whether
sampling or filtering was applied, which traffic class or forwarding
treatment was observed, and whether the value is still fresh enough
for the intended decision.
3. Terminology
Observation Point: The location and logical boundary at which a
measurement is made. Examples include ingress interface, egress
interface, device internal counter, endpoint software boundary,
RDMA operation boundary, fabric boundary, or service contact
instance boundary.
Measurement Domain: The set of conditions under which a telemetry
value is meaningful. This can include observation point, traffic
class, flow scope, encapsulation state, sampling rule, aggregation
interval, and metric source.
Comparability Domain: The set of values with which a telemetry value
may be compared without changing its meaning. Values outside this
domain may still be useful, but the consumer needs a rule for
interpreting the difference.
Freshness Context: Information that allows a consumer to determine
when a metric was measured, when it was exported, and whether it
is still valid for the decision being made.
Aggregation Scope: The set of packets, flows, traffic classes,
interfaces, service instances, or time intervals included in a
reported value.
Series Identity: The set of fields that identifies one continuous
measurement series. A change in exporter, observation point,
sampling rule, counter state, traffic class, aggregation interval,
or timestamp basis can create a different series.
4. Motivating Cases
4.1. ECN and Congestion Telemetry
ECN telemetry can be exported as packet counts, byte counts, marking
ratios, codepoint observations, or flow records. A collector needs
to know whether the value was observed before or after remarking,
decapsulation, encapsulation, load balancing, or traffic class
mapping. It also needs to know whether the reported series includes
all packets for the flow, sampled packets, one direction of the flow,
or both directions.
If those conditions are not explicit, an ECN marking rate can be
compared with the wrong denominator or correlated with the wrong
traffic direction.
Example: an exporter reports a CE packet counter for a flow at
ingress, while another exporter reports a CE packet counter for the
same apparent flow at egress. If traffic was remarked, tunneled,
sampled, or load balanced between the two points, the two counters do
not by themselves describe the same series. A collector needs the
observation point, direction, traffic class, sampling rule, packet or
byte denominator, and interval before it can compute or compare a
marking rate.
4.2. Queue and Delay Measurements
Queue and delay measurements are sensitive to the observation
boundary. An endpoint boundary measurement can include application,
transport, and completion queue behavior. A network boundary
measurement can include fabric delay but not endpoint processing. A
device internal queue measurement can expose a useful signal, but it
may be implementation specific and not directly comparable with
another device.
The exported value therefore needs to carry enough context for a
consumer to know whether it is comparing endpoint delay, fabric
delay, queue occupancy, or a combined value.
Example: one system reports latency from an application request to an
application response. Another system reports delay between two
fabric boundary points. Both values can be useful, but they should
not be ranked as if they were the same metric. The first value
includes endpoint and service processing. The second value describes
network forwarding within a selected boundary.
4.3. AI Network Fabric Measurements
AI network fabric measurements often combine workload behavior,
endpoint behavior, and fabric behavior. Training and inference
traffic can use different message sizes, burst patterns, traffic
classes, congestion control parameters, and pause behavior. A result
that reports latency, ECN marking, PFC events, queue depth, or
throughput is not comparable unless the observation point and traffic
conditions are known.
Controlled benchmarking work can define these conditions for a lab
test. Operational telemetry needs a related but separate rule: the
exported measurement should state the observation context and the
comparability domain so that values collected from live systems are
not mixed incorrectly.
Example: a controller compares queue depth from one switch port with
a completion latency value from an endpoint. The two values may both
indicate congestion pressure, but they do not measure the same part
of the system. A consumer needs the observation boundary before it
can decide whether the values support the same conclusion.
4.4. Flow Correlation Across Observation Points
Flow identifiers and correlation keys can be reused or scoped by an
exporter, observation domain, interface, tunnel, service instance, or
measurement session. A collector needs to know the uniqueness scope
before it joins records from different sources.
A common numeric identifier does not by itself prove that two records
describe the same flow or the same measurement event. The exported
context must define the joining rule.
Example: two exporters use the same flow identifier value. The value
is unique only within each exporter and observation domain. Joining
the records on the numeric identifier alone can merge unrelated
measurements. The join rule needs to include the exporter,
observation domain, and any measurement session or namespace needed
to make the identifier unique.
4.5. Metric Freshness for Traffic Steering
Traffic steering systems can combine service metrics and network
metrics. A path selector might use service load, network delay,
packet loss, and reachability to select a service contact instance or
egress forwarder. The selected result can be wrong if these inputs
were measured at different times or have different validity periods.
Example: a service load metric says that a service instance is
lightly loaded, but the value was collected before a sudden workload
change. At the same time, a fresh network metric says that the
network path is healthy. Combining the stale service metric with the
fresh network metric can steer traffic to an overloaded instance. A
consumer needs freshness context and a validity rule before using
those values in one decision.
Freshness is also part of stability. If steering decisions react to
every small metric change, the system can create oscillation. A
specification that expects automated use should identify whether hold
time, damping, thresholding, or another stability method is expected.
5. Comparability Requirements
Specifications that define or use operational telemetry for IP
performance measurement should state the following information when
the values are expected to be compared, aggregated, or used by
automated systems.
REQ1: The specification should identify the Observation Point for
each metric or metric series.
REQ2: The specification should identify the traffic class,
forwarding treatment, encapsulation state, and direction of
traffic to which the value applies, when these affect the
interpretation of the metric.
REQ3: The specification should identify the Aggregation Scope of the
value, including the time interval, packet set, flow set,
interface set, service instance set, or traffic class set
represented by the metric.
REQ4: The specification should state the Comparability Domain for
the value. If values from different observation points, traffic
classes, or aggregation scopes are not directly comparable, that
condition should be stated.
REQ5: The specification should define how a consumer detects a new
measurement series when the observation point, exporter, sampling
rule, counter state, traffic class, or aggregation interval
changes.
REQ6: The specification should provide Freshness Context when a
measurement can be used by a controller, path selector, assurance
system, or other automated consumer. This can include timestamp,
sequence number, validity interval, or update reason.
REQ7: The specification should state how missing records, counter
resets, exporter restarts, and stale values are represented. A
missing delta record should not be silently interpreted as zero
traffic unless the specification explicitly defines that behavior.
REQ8: The specification should define the uniqueness scope and join
rule for any identifier used to correlate measurements across
observation points or exporters.
REQ9: The specification should state whether the measurement method
is active, passive, or hybrid, when that affects interpretation of
the value. If the method changes the packet stream, observes an
existing packet stream, or adds in situ data to packets, that fact
should be visible to the consumer.
REQ10: The specification should identify the timestamp basis and
clock context used by a metric when timing affects comparison.
Values measured using different timestamp formats, clock domains,
or synchronization assumptions should not be treated as directly
comparable unless the specification defines the conversion rule.
6. Relationship to Existing Work
RFC 2330 defines the general IPPM goal of concrete, well defined,
repeatable, and useful metrics. This document does not alter that
framework. It applies the same discipline to operational telemetry
values that are consumed after export by collectors and automated
systems.
RFC 7799 defines active, passive, and hybrid methods. This document
uses that distinction because the method affects the meaning of a
value. A metric derived from generated test traffic, a metric based
only on observed production packets, and a metric derived from in
situ data placed in production packets can have different assumptions
about packet selection, packet modification, and measurement points.
RFC 7011 and RFC 7012 define IPFIX observation points, observation
domains, scoped fields, counters, and flow records. This document is
consistent with that model. It asks specifications that use exported
measurements to make the observation and scope rules visible when a
consumer is expected to compare or aggregate values.
RFC 5475 describes packet selection, sampling, filtering, and packet
streams. Sampling and filtering are part of the measurement domain.
A metric collected from sampled packets is not automatically
comparable with a metric collected from all packets unless the
specification defines the relationship.
RFC 8321 and related Alternate Marking work show that counters from
multiple measurement points are comparable only when they refer to
the same packet block or marking period. This document generalizes
that lesson: a consumer needs a common aggregation scope before it
compares values from different points.
RFC 9197 defines IOAM data fields collected within a limited domain.
IOAM namespaces provide context for interpreting node identifiers,
timestamp formats, queue data, and other fields. This document does
not change IOAM. It uses IOAM as an example of why context and
namespace are part of measurement interpretation.
Benchmarking methodology work can define controlled test procedures,
reporting requirements, and benchmark conditions. This document is
concerned with operational telemetry used outside a single controlled
test. The two topics are related because both need clear observation
points and comparability rules, but the deployment assumptions are
different.
This document is also related to telemetry identifier scoping work.
Identifier scoping addresses whether a value uniquely identifies the
same object or observation. This document focuses on whether
measurement values may be compared or aggregated. Both properties
are needed by collectors and automated consumers.
Existing IPPM metrics and measurement protocols remain unchanged by
this document. The requirements in this document are intended to
help future specifications state the context needed to interpret
those metrics in operational telemetry systems.
7. Security Considerations
Missing or incorrect comparability context can lead to incorrect
operational decisions. A controller could steer traffic away from a
healthy path, select an overloaded service instance, or suppress a
valid alarm if it compares stale or mismatched measurements.
Telemetry context can also expose sensitive operational information,
including topology, traffic class use, service placement, load, and
congestion state. Specifications using the requirements in this
document should consider confidentiality, integrity, and access
control for both the measurement value and its context.
8. IANA Considerations
This document has no IANA actions.
9. Acknowledgements
The motivation for this document came from discussions about IPPM,
BMWG, IPFIX, and operational telemetry models where measurement
values need explicit scope before they can be compared.
10. References
10.1. Informative References
[RFC2330] Paxson, V., Almes, G., Mahdavi, J., and M. Mathis,
"Framework for IP Performance Metrics", RFC 2330, May
1998.
[RFC5475] Zseby, T., Molina, M., Duffield, N., Niccolini, S., and F.
Raspall, "Sampling and Filtering Techniques for IP Packet
Selection", RFC 5475, March 2009.
[RFC7011] Claise, B., Trammell, B., and P. Aitken,
"Specification of the IP Flow Information Export Protocol
for the Exchange of Flow Information", RFC 7011,
September 2013.
[RFC7012] Claise, B. and B. Trammell, "Information Model for IP Flow
Information Export", RFC 7012, September 2013.
[RFC7799] Morton, A., "Active and Passive Metrics and Methods",
RFC 7799, May 2016.
[RFC8321] Fioccola, G., Ed., Cociglio, M., Sapio, A., Sisto, R., and
T. Mizrahi, "Alternate Marking Method for Passive and
Hybrid Performance Monitoring", RFC 8321, January 2018.
[RFC9197] Brockners, F., Bhandari, S., and T. Mizrahi, "Data Fields
for In Situ Operations, Administration, and Maintenance",
RFC 9197, May 2022.
[I-D.ietf-ippm-on-path-telemetry-yang]
Fioccola, G., Zhou, T., Zhu, Y., Zhang, W., and K. Zhu,
"On-Path Telemetry YANG Data Model",
draft-ietf-ippm-on-path-telemetry-yang-06, August 2026.
[I-D.ietf-ippm-alt-mark-yang]
Graf, T., Wang, M., Fioccola, G., Zhou, T., and X. Min,
"A YANG Data Model for the Alternate-Marking Method",
draft-ietf-ippm-alt-mark-yang-06, August 2026.
[I-D.ietf-cats-framework]
Li, C., Du, Z., Boucadair, M., Contreras, L., and J.
Drake, "A Framework for Computing-Aware Traffic Steering",
draft-ietf-cats-framework-24, April 2026.
Author's Address
Saumya Dikshit
HPE
Email: saumya.dikshit@hpe.com
Dikshit Expires 6 May 2027 [Page 9]