BPM L. Liu
Internet-Draft Zhongguancun Laboratory
Intended status: Informational N. Geng
Expires: 6 May 2027 Huawei Technologies
Y. Chen
Zhongguancun Laboratory
2 November 2026
Benchmarking Methodology for Route Origin Validation (ROV)
draft-lg-bpm-benchmarking-methodology-for-rov-00
Abstract
This document defines a benchmarking methodology for routers that
implement Route Origin Validation (ROV). The methodology focuses on
device-level behavior, including processing of validated Route Origin
Authorization (ROA) payload (VRP) updates, the interaction between
ROV and BGP, resource utilization, and the scalability of ROV under
varying operational conditions. The procedures described here follow
the principles and constraints of the Benchmarking Methodology
Working Group (BMWG) and are intended to produce repeatable and
comparable results across implementations.
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.
Liu, et al. Expires 6 May 2027 [Page 1]
Internet-Draft ROV Benchmarking November 2026
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.
Table of Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1. Requirements Language . . . . . . . . . . . . . . . . . . 4
2. Scope and Goals . . . . . . . . . . . . . . . . . . . . . . . 4
3. Terminology . . . . . . . . . . . . . . . . . . . . . . . . . 5
4. Test Setup and Laboratory Environment . . . . . . . . . . . . 6
4.1. Test Topology . . . . . . . . . . . . . . . . . . . . . . 6
4.2. DUT Configuration Requirements . . . . . . . . . . . . . 7
4.3. RTR Data Source Emulation . . . . . . . . . . . . . . . . 8
4.4. BGP Traffic Generation Requirements . . . . . . . . . . . 8
4.5. Traffic Profile Parameters . . . . . . . . . . . . . . . 8
5. Benchmarking Methodology . . . . . . . . . . . . . . . . . . 9
5.1. General Considerations . . . . . . . . . . . . . . . . . 9
5.2. Test Control and Input Conditions . . . . . . . . . . . . 9
5.3. Metrics and Measurements . . . . . . . . . . . . . . . . 10
6. Benchmark Tests . . . . . . . . . . . . . . . . . . . . . . . 11
6.1. ROV Update Processing Latency . . . . . . . . . . . . . . 12
6.2. ROV Validation Latency . . . . . . . . . . . . . . . . . 13
6.3. BGP Convergence with ROV Enabled . . . . . . . . . . . . 14
6.3.1. Case A: VRP-triggered BGP convergence . . . . . . . . 14
6.3.2. Case B: BGP-triggered convergence with and without
ROV . . . . . . . . . . . . . . . . . . . . . . . . . 16
6.4. VRP Scalability Tests . . . . . . . . . . . . . . . . . . 17
6.5. VRP Churn and Stress Tests . . . . . . . . . . . . . . . 18
6.6. Resource Utilization . . . . . . . . . . . . . . . . . . 18
6.7. RTR Session Behavior Tests . . . . . . . . . . . . . . . 19
7. Reporting Requirements . . . . . . . . . . . . . . . . . . . 21
8. Security Considerations . . . . . . . . . . . . . . . . . . . 23
9. Implementation Status . . . . . . . . . . . . . . . . . . . . 23
10. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 24
11. References . . . . . . . . . . . . . . . . . . . . . . . . . 24
11.1. Normative References . . . . . . . . . . . . . . . . . . 24
11.2. Informative References . . . . . . . . . . . . . . . . . 25
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . 26
Appendix A. Summary of Changes . . . . . . . . . . . . . . . . . 26
Appendix A.1. Changes from
draft-lg-bmwg-benchmarking-methodology-for-rov-03 to
draft-lg-bpm-benchmarking-methodology-for-rov-00 . . . 26
Liu, et al. Expires 6 May 2027 [Page 2]
Internet-Draft ROV Benchmarking November 2026
Appendix A.2. Changes from -02 to -03 . . . . . . . . . . . . 27
Appendix A.3. Changes from -01 to -02 . . . . . . . . . . . . 27
Appendix A.4. Changes from -00 to -01 . . . . . . . . . . . . 28
Appendix B. Implementation Details . . . . . . . . . . . . . . . 28
Appendix B.1. Test Environment . . . . . . . . . . . . . . . . 28
Appendix B.2. BGP and VRP Data . . . . . . . . . . . . . . . . 29
Appendix B.3. Measurement Methods . . . . . . . . . . . . . . 31
Appendix B.4. General Conventions . . . . . . . . . . . . . . 32
Appendix B.5. ROV Update Processing Latency . . . . . . . . . 33
Appendix B.6. ROV Validation Latency . . . . . . . . . . . . . 33
Appendix B.7. BGP Convergence with ROV Enabled . . . . . . . . 34
Appendix B.7.1. Case A: VRP-triggered BGP convergence . . . 34
Appendix B.7.2. Case B: BGP-triggered convergence with and
without ROV . . . . . . . . . . . . . . . . . . . . . 35
Appendix B.8. VRP Scalability Tests . . . . . . . . . . . . . 37
Appendix B.9. VRP Churn and Stress Tests . . . . . . . . . . . 38
Appendix B.10. Resource Utilization . . . . . . . . . . . . . 43
Appendix B.11. RTR Session Behavior Tests . . . . . . . . . . 44
Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 47
1. Introduction
Route Origin Validation (ROV), as specified in [RFC6811], allows
routers to use validated Route Origin Authorization (ROA)
information, which is distributed via the RPKI-to-Router (RTR)
protocol defined in [RFC8210], to classify BGP routes as Valid,
Invalid, or NotFound. Deployments of ROV continue to increase across
networks, and router vendors have implemented ROV processing as part
of their control-plane functions.
While operational experience is growing, there is currently no
standardized methodology for measuring the performance impact and
behavioral characteristics of ROV on routing devices. As with other
protocol features evaluated by the Benchmarking Methodology Working
Group (BMWG), a consistent and repeatable test framework is essential
for:
* Comparing router implementations,
* Evaluating scalability under controlled conditions,
* Characterizing the control-plane costs of ROV processing, and
* Understanding how ROV influences BGP convergence and routing
stability.
Liu, et al. Expires 6 May 2027 [Page 3]
Internet-Draft ROV Benchmarking November 2026
This document defines a benchmarking methodology for routers that
implement ROV, which builds upon the foundational benchmarking
principles defined in [RFC1242], [RFC2285], [RFC2544], [RFC2889], and
[RFC3918]. The methodology focuses on the Device Under Test (DUT)
and uses controlled, reproducible inputs to isolate the effects of
ROV from external dependencies. In particular, the benchmarking
framework assumes the presence of an RTR update source, which may be
an RPKI Cache Server or an RTR traffic generator capable of
delivering synthetic Validated ROA Payloads (VRPs).
The objective of this document is to define a set of metrics and
procedures to quantify:
* The latency of ROV state updates within the DUT,
* The impact of ROV on BGP control-plane performance,
* The scalability of ROV processing under varying VRP and BGP table
sizes, and
* The resource utilization associated with enabling ROV.
By providing a consistent framework, this document enables vendors,
operators, and researchers to evaluate ROV functionality under
controlled and repeatable conditions, improving understanding of
implementation performance and supporting informed deployment
decisions.
1.1. Requirements Language
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT",
"SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and
"OPTIONAL" in this document are to be interpreted as described in
BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all
capitals, as shown here.
2. Scope and Goals
This document specifies a laboratory-based benchmarking methodology
for evaluating the performance of router implementations of ROV as
defined in [RFC6811]. The scope of this benchmarking methodology
includes:
* *ROV processing performance*: Measurement of the time and
resources required for a router to process VRP updates received
via the RTR protocol.
Liu, et al. Expires 6 May 2027 [Page 4]
Internet-Draft ROV Benchmarking November 2026
* *Impact on BGP control-plane performance*: Quantification of how
enabling ROV affects BGP convergence times and routing table
stability.
* *Scalability under controlled conditions*: Evaluation of the
router's ability to handle large VRP sets, rapid VRP churn, and
BGP updates influenced by ROV.
* *Resource utilization*: Measurement of system CPU utilization,
system memory consumption, and relevant control-plane process load
associated with ROV processing.
The goals of this document are:
* To define a repeatable, controlled methodology for benchmarking
ROV-enabled routers.
* To provide standardized metrics that allow for comparison across
implementations.
3. Terminology
The terminology used in this document follows the conventions of
[RFC1242], [RFC2285], and subsequent BMWG publications. The
following terms are used with specific meanings in the context of ROV
benchmarking.
Route Origin Validation (ROV): A procedure defined in [RFC6811] that
compares the origin AS of a BGP announcement with the set of
authorized origins derived from validated ROA objects. ROV results
in one of three states: Valid, Invalid, or NotFound.
Validated ROA Payload (VRP): The processed output from a relying
party containing prefix-origin pairs that routers use for ROV
decisions. VRPs are transported via the RTR protocol.
RPKI-to-Router Session: A protocol session between a router and an
RPKI Cache Server. In benchmarking, RTR sessions may be emulated or
generated using traffic/test tools to deliver synthetic VRP updates.
ROV Update Processing Latency: The time from when a router receives
new VRP data (via RTR) until the updated ROV state is reflected in
the router's local Routing Information Base (RIB) or implemented in
routing decisions.
VRP-Triggered Revalidation Latency: The time interval between
completion of VRP installation and the moment all affected prefixes
have updated validation states.
Liu, et al. Expires 6 May 2027 [Page 5]
Internet-Draft ROV Benchmarking November 2026
BGP-Triggered ROV Validation Latency: The time interval between
receipt of a BGP UPDATE message and completion of the ROV validation
procedure for that route.
BGP Convergence Time: The time required for the router's control
plane to process BGP updates and reach a stable routing state, while
ROV validation is active.
Resource Utilization: System CPU utilization, system memory
consumption, and, when observable, per-process utilization of the ROV
process and the BGP or routing process while the router performs ROV-
related tasks, including processing of VRP updates and applying ROV
policy.
ROV Churn: A burst of VRP changes (e.g., many ROA additions or
withdrawals) that may trigger significant revalidation and BGP
recalculation, which is used in stress tests.
ROV Scalability Limit: The maximum number of VRPs, RTR sessions, or
ROV-triggered BGP changes that the router can process while
maintaining normal operational performance.
4. Test Setup and Laboratory Environment
This section describes the required test topology, equipment, DUT
configuration, RPKI data emulation, and traffic generation
conditions. The goal of the test environment is to isolate the DUT
and subject it to clearly defined RPKI-RTR and BGP tests, while
providing accurate timing and state measurements.
4.1. Test Topology
+-------------------+ RTR +----------------------+
| RTR Emulator |---------->| DUT |
|(RTR Update Source)| | (ROV Enabled) |
+-------------------+ +----------------------+
/\ /\
| | Data-plane Traffic
BGP | +-----------------+
+---------------------+ | | Tester |
|BGP Traffic Generator|-----------+ |(Data-plane Load)|
+---------------------+ +-----------------+
Figure 1: The test topology for ROV benchmarking.
The test topology consists of four primary components: the DUT, an
RPKI-RTR update source, a BGP traffic generator, and a tester for
generating data-plane traffic load. The DUT is a router equipped
Liu, et al. Expires 6 May 2027 [Page 6]
Internet-Draft ROV Benchmarking November 2026
with ROV capabilities, supporting the RPKI-RTR protocol and applying
ROV policies to received BGP routes. The RPKI-RTR update source may
be either a real RPKI cache implementation running in isolated mode
or a dedicated emulator capable of producing arbitrary VRP sets and
update patterns. This RTR source connects directly to the DUT using
the RTR protocol and provides controlled VRP updates, including
serial increments, cache resets, and bursty or delayed update
sequences.
The BGP traffic generator establishes one or more BGP peering
sessions with the DUT and is responsible for delivering a full
routing table together with controlled withdrawal or re-announcement
events. Because IPv4 and IPv6 tables differ in scale and may
exercise different implementation paths, the test setup must state
the number of IPv4 routes and the number of IPv6 routes separately.
A test MAY use a mixed baseline table (for example, 1,000,000 IPv4
routes and 250,000 IPv6 routes) or MAY benchmark each address family
identifier (AFI) separately. In either case, the chosen route counts
and AFIs under test must remain fixed across repeated runs for the
same condition. The generator should be capable of presenting both
stable baseline routing conditions and timed ROV-affected prefixes
whose validation status will change in response to VRP updates. A
tester is connected to the DUT to introduce controlled data-plane
load during benchmarking. When present, the tester should generate
stable and deterministic traffic loads so that the impact of
forwarding load on ROV processing can be evaluated. When data-plane
load is applied, its rate, frame size, traffic pattern, and address-
selection rules needs be documented in the test report.
4.2. DUT Configuration Requirements
The DUT must be configured with ROV enabled on all BGP sessions
receiving test routes. The router needs to establish a stable and
fully functional RPKI-RTR session with the RTR emulator. To ensure
that performance results are attributable solely to ROV behavior, all
non-essential features on the DUT, such as additional routing
protocols, unnecessary telemetry mechanisms, and unused services,
should be disabled. Logging related to ROV may remain enabled for
debugging purposes but can be rate-limited to avoid skewing CPU
measurements or affecting test repeatability. All system parameters
relevant to routing performance, such as multipath behavior or
maximum-prefix limits, is suggested to be documented prior to
testing.
Liu, et al. Expires 6 May 2027 [Page 7]
Internet-Draft ROV Benchmarking November 2026
4.3. RTR Data Source Emulation
The RTR emulator is used to generate synthetic VRP data sets with
user-defined characteristics. This includes the ability to create
arbitrary combinations of prefixes and ASNs, overlapping VRPs,
conflicting VRPs, and other edge cases relevant to validation logic.
The VRP datasets should mimic realistic global distributions where
appropriate, but can also support scaling tests where VRP volumes are
substantially higher than today's norm. The data source needs to be
able to further support generating controlled bursts of VRP updates,
ranging from 100 to 10,000 VRP changes per second, and can allow for
both additive updates and withdrawals.
4.4. BGP Traffic Generation Requirements
The BGP traffic generator should be able to present the DUT with a
stable baseline routing table prior to initiating any benchmark.
This ensures that the DUT begins each test run in a known, converged
state with predictable CPU and memory utilization. The generator can
also provide a set of ROV-affected prefixes whose origin AS can be
manipulated in concert with VRP updates from the RTR emulator. These
prefixes should span a range of prefix lengths and originate from
diverse ASes to reflect realistic routing conditions. The traffic
generator can support deterministic convergence triggers, such as the
precise injection of BGP updates following a VRP change or the
simultaneous application of both BGP and VRP events.
4.5. Traffic Profile Parameters
When data-plane traffic is used, the following parameters should be
specified:
* Fixed frame size used for the measurement. For convergence
measurements, a small fixed packet size SHOULD be used to improve
time resolution. A 128-byte packet at Layer 3 is one practical
choice. Other fixed sizes MAY be used when required by the
traffic generator or encapsulation overhead, but the selected size
needs to be documented.
* Traffic rate in packets per second (PPS). For convergence
measurements, the traffic should use a constant rate so that
packet arrival times map directly to time resolution. Burst
traffic MAY be used in stress scenarios, but such tests must be
reported separately from constant-rate convergence tests.
* Traffic pattern. For convergence measurements, constant-rate
traffic is RECOMMENDED.
Liu, et al. Expires 6 May 2027 [Page 8]
Internet-Draft ROV Benchmarking November 2026
* Source and destination IP address selection rules. When the setup
sends one stream per tested prefix, the destination address for
that stream should be selected from within the tested prefix. The
first usable address is one valid example. The selection rule
needs to be documented and applied consistently across runs.
* Whether traffic matches ROV-affected prefixes.
Each frame size and PPS combination should be reported separately.
5. Benchmarking Methodology
This section describes the general methodology for benchmarking ROV
behavior on a DUT. The goal is to ensure that all tests are
repeatable, comparable across different environments, and
representative of realistic deployment conditions. The methodology
defines how to establish a controlled and stable test environment,
how to specify and vary input conditions, and how to measure key
performance metrics associated with ROV processing.
5.1. General Considerations
Before any measurements are taken, the DUT must reach a well-defined
steady state in which the RPKI-RTR session is fully established, the
VRP set has been completely synchronized, and the BGP control plane
has converged. A warm-up period is recommended to eliminate any
cold-start effects that could bias measurement results.
All sources of measurement noise should be avoided. Features such as
logging, real-time telemetry export, or periodic background tasks can
interfere with timing-sensitive measurements; therefore, such
features should be disabled or rate-limited during benchmarking. CPU
clock scaling, thermal throttling, or other variable-performance
modes should be minimized if the test setup allows it.
5.2. Test Control and Input Conditions
Accurate benchmarking depends on precise control of the input
conditions applied to the DUT. All tests should begin from a
consistent baseline consisting of:
* A predefined ROA and VRP population, including the total number of
ROAs and the total number of resulting VRPs. When synthetic ROAs
are used to generate VRPs, the relationship between ROA count and
VRP count must be stated.
Liu, et al. Expires 6 May 2027 [Page 9]
Internet-Draft ROV Benchmarking November 2026
* A stable and realistic baseline BGP RIB-in with IPv4 and IPv6
route counts documented separately (for example, 1,000,000 IPv4
routes and 250,000 IPv6 routes in a mixed table, or equivalent
AFI-specific baselines).
* A defined coverage level of the VRP dataset relative to the
announced routing space. Coverage SHOULD be reported as the
percentage of announced prefixes, or the percentage of announced
address space, for which at least one matching VRP exists. The
selected coverage definition neds to be stated and applied
consistently across runs.
* Defined structural properties of the VRP dataset, including
overlap characteristics. At minimum, the test should specify the
fraction or probability that a VRP is covered by another VRP, for
example where a more-specific VRP falls within the prefix range of
a less-specific VRP. Tests MAY additionally report the
distribution of covering depth, prefix lengths, and authorized
origin AS diversity when those properties are relevant to the
implementation under test.
From this baseline, input variables may be modified to stress
different aspects of ROV behavior. These variables include the VRP
churn rate, ranging from steady incremental updates to high-intensity
bursts, the coverage level of the VRP dataset, the overlap
probability among VRPs, and the type of RPKI-RTR updates provided to
the DUT, such as incremental updates versus full-table refreshes.
Each of these conditions may trigger different processing strategies
within the DUT, and therefore needs to be explicitly controlled and
documented.
5.3. Metrics and Measurements
Benchmarking ROV behavior requires collecting quantitative
performance metrics that reflect how the DUT processes validation
information and incorporates it into the BGP decision process.
Therefore, this document proposes key performance metrics including
ROV update processing latency, ROV validation latency, BGP
convergence time, VRP storage size, system CPU and memory
utilization, per-process utilization where available, and ROV state
rebuild time.
*ROV update processing latency* measures the time from receipt of an
RTR update (incremental or full) until the DUT has fully updated its
internal validation state. This metric captures the efficiency of
ROV data structures and algorithms.
Liu, et al. Expires 6 May 2027 [Page 10]
Internet-Draft ROV Benchmarking November 2026
*ROV validation latency* measures the time interval between a
router's receipt of a BGP UPDATE message that contains a new or
changed route, and the completion of the ROV procedure for that
route, producing a validation state of Valid, Invalid, or NotFound.
This metric isolates the internal validation step, excluding the
larger BGP convergence process, and provides insight into the
responsiveness of the DUT's validation engine.
*BGP convergence time* with ROV enabled measures how long the DUT
takes to converge on BGP prefixes whose validation states change due
to VRP updates. This reflects the real operational behavior of ROV
as it interacts with the control plane.
The *VRP storage size* inside the DUT should also be recorded to
evaluate the scalability of the implementation when operating with
large VRP datasets. Alongside this, *resource utilization* SHOULD be
monitored to identify performance limits or resource-intensive
operations triggered by ROV. At minimum, the benchmark should
collect system-level CPU utilization and system-level memory
consumption. When the DUT exposes process-level counters, the
benchmark should also collect utilization for the ROV process and for
the BGP process or main routing process.
A recovery-related measurement, *ROV state rebuild time* after RTR
session reset, quantifies the time needed for the DUT to re-establish
a complete and correct ROV validation state after an RTR session
reset or cache outage. This metric reflects robustness and recovery
behavior under fault or restart scenarios.
Finally, the DUT should be evaluated under high-pressure scenarios by
measuring its behavior when processing VRP bursts, such as surges of
100-10,000 VRPs per second. This measurement reveals whether the
implementation can sustain abrupt workload increases without dropping
updates, stalling, or entering unstable states.
6. Benchmark Tests
This section defines the individual benchmark tests used to evaluate
the performance and behavior of a DUT implementing ROV. Each test
focuses on a specific aspect of the ROV processing pipeline,
including VRP ingestion, validation, interaction with BGP,
scalability limits, and robustness under stress and failure
conditions. All tests assume the laboratory setup and input
conditions described previously.
Liu, et al. Expires 6 May 2027 [Page 11]
Internet-Draft ROV Benchmarking November 2026
6.1. ROV Update Processing Latency
*Objective*: Measure the latency from the arrival of an RTR PDU until
the new VRP information is installed in the DUT's internal ROV
tables.
The *test procedures* for ROV update processing latency are listed
below:
1. Prepare baseline state
* Establish RTR session between DUT and the RTR emulator.
* Preload DUT with a selected baseline VRP size (e.g., 100k
VRPs).
* Ensure BGP is fully converged.
2. Inject controlled RTR update
* From the emulator, send a controlled incremental update. The
test must state whether the update affects a single VRP or a
batch of multiple VRPs.
* For each changed VRP, document whether the operation is an
addition, withdrawal, or modification.
* Document the expected validation-state change triggered by the
update, such as Valid-to-Invalid due to origin-AS mismatch,
Valid-to-Invalid due to prefix-length violation, Invalid-to-
Valid, Valid-to-NotFound, Invalid-to-NotFound, or NotFound-to-
Valid.
* Alternatively, for full-refresh tests, send a full VRP set
replacement PDU sequence and document the number and type of
VRP changes contained in the replacement.
3. Timestamp PDU transmission
* Record the exact moment the first update PDU is sent.
4. Monitor DUT internal state
* Use device instrumentation (API, CLI, or telemetry) to detect
the exact moment the VRP table reflects the update.
* Confirm the VRP entry has been added, removed, or modified as
expected.
Liu, et al. Expires 6 May 2027 [Page 12]
Internet-Draft ROV Benchmarking November 2026
5. Calculate latency
* The latency is the time difference between the moment the RTR
PDU is sent and the moment the VRP is applied on the DUT. For
batch updates, the report should distinguish first-VRP
installation latency, last-VRP installation latency, and
aggregate batch processing time when the DUT exposes
sufficient instrumentation.
6. Repeat for multiple VRP table sizes
* E.g., 50k, 100k, 500k, and 1M VRPs.
7. Repeat no less than 10 independent runs per condition
* Compute mean and standard deviation.
6.2. ROV Validation Latency
*Objective*: Measure how long the DUT takes to apply updated VRPs to
the validation states of affected BGP prefixes.
The *test procedures* for ROV validation latency are listed below:
1. Establish baseline
* Load the baseline BGP table with IPv4 and IPv6 route counts
stated explicitly (for example, 1,000,000 IPv4 routes and
250,000 IPv6 routes, or an AFI-specific baseline).
* Ensure all prefixes have a known baseline validation state.
2. Select a controlled prefix set
* Pick a set of prefixes (e.g., 1,000) whose origin AS is tied
to specific VRPs.
* Control and report the distribution of affected prefixes
across the validation structure. At minimum, include one
scenario in which affected prefixes are localized to the same
region of the validation structure (for example, sharing a
common covering prefix, trie branch, or address block) and one
scenario in which affected prefixes are distributed across
different regions.
* The locality rule, address-family mix, prefix-length
distribution, and number of affected prefixes per region needs
to be documented.
Liu, et al. Expires 6 May 2027 [Page 13]
Internet-Draft ROV Benchmarking November 2026
3. Trigger validation update
* Modify VRPs so that these prefixes change validation state
(Valid->Invalid or Invalid->Valid).
* The test must document whether each validation change is
caused by an origin-AS mismatch, a prefix-length violation,
loss of covering VRP, addition of a covering VRP, or another
explicitly stated condition.
4. Timestamp VRP installation completion
* As measured in Section 6.1.
5. Monitor DUT validation table
* Continuously query validation state for selected prefixes.
* Note the timestamp when all prefixes reflect the updated
state.
6. Compute latency
* The validation latency is the time difference between the
moment the VRP installation is completed and the moment all
affected validation states have been updated.
7. Repeat with varying set sizes
* E.g., 10 prefixes, 100 prefixes, 1,000 prefixes.
6.3. BGP Convergence with ROV Enabled
*Objective*: Measure BGP convergence time for routes impacted by ROV
state changes, and compare ROV-triggered convergence with BGP-only
convergence.
This test includes two cases. The first case measures convergence
caused by VRP changes delivered through RTR, where the intent is to
trigger ROV revalidation and then observe the resulting BGP behavior.
The second case measures the propagation and convergence time of the
same BGP routing event with ROV enabled and with ROV disabled.
6.3.1. Case A: VRP-triggered BGP convergence
The *test procedures* for VRP-triggered BGP convergence are listed
below:
Liu, et al. Expires 6 May 2027 [Page 14]
Internet-Draft ROV Benchmarking November 2026
1. Prepare baseline
* Establish full-table BGP adjacency.
* Enable ROV on the DUT.
* Ensure stable initial convergence.
2. Select test prefixes
* Choose prefixes that will transition validation state once VRP
updates are applied, such as Valid-to-Invalid or Invalid-to-
Valid.
* Document the cause of the validation change, such as origin-AS
mismatch or prefix-length violation.
3. Trigger VRP state change
* Send VRP modifications via RTR. The purpose of this step is
to trigger convergence through VRP changes delivered by RTR
and then observe the BGP control-plane behavior that follows
from the changed validation state.
4. Monitor BGP behavior
* Observe best-path selection changes.
* Timestamp withdrawal, replacement, or re-installation of
affected prefixes.
5. Measure convergence
* End-to-end convergence timer: This timer should start when the
RTR emulator sends the first relevant VRP update and should
end when both the BGP RIB and Forwarding Information Base
(FIB) reach stable state. This metric captures RTR delivery,
VRP installation, ROV revalidation, and the resulting BGP
convergence.
* Post-installation BGP convergence timer: This timer should
start when the DUT completes installation of the relevant VRP
update and should end when both the BGP RIB and FIB reach
stable state. This metric isolates BGP convergence after VRP
installation.
* If only one timer is reported, the report needs to state which
timer definition is used.
Liu, et al. Expires 6 May 2027 [Page 15]
Internet-Draft ROV Benchmarking November 2026
6. Record:
* Time to withdraw or replace routes that become Invalid.
* Time to install or select routes that become Valid.
* Time until new best paths stabilize.
6.3.2. Case B: BGP-triggered convergence with and without ROV
The *test procedures* for comparing BGP-triggered convergence are
listed below:
1. Prepare baseline conditions
* Prepare two otherwise identical baseline conditions: one with
ROV enabled and one with ROV disabled.
* Ensure that the baseline BGP table, VRP dataset, DUT
configuration, data-plane traffic conditions, and measurement
instrumentation are identical unless explicitly varied by the
test.
2. Inject the BGP event
* Apply the same BGP event in both conditions.
* Example events include announcing a new route, withdrawing a
route, changing an origin AS, or changing BGP attributes that
affect best-path selection.
3. Control affected prefixes and distribution
* Ensure that the route count, affected prefix set, prefix-
length distribution, AFI mix, and structural distribution of
affected prefixes are identical between the ROV-enabled and
ROV-disabled runs.
* If data-plane traffic is used, the traffic rate, frame size,
destination-prefix mapping, and address-selection rules must
also remain identical.
4. Measure convergence time
* Measure the time from BGP event injection to the point at
which both the BGP RIB and FIB reach stable state.
Liu, et al. Expires 6 May 2027 [Page 16]
Internet-Draft ROV Benchmarking November 2026
* The timer start and timer end conditions must be identical for
the ROV-enabled and ROV-disabled runs.
5. Record and compare results
* Record the absolute convergence time for both runs.
* Report the relative difference introduced by enabling ROV,
including minimum, average, maximum, and at least P95 values
across repeated runs.
6.4. VRP Scalability Tests
*Objective*: Evaluate DUT performance with varying VRP table sizes.
The *test procedures* for VRP scalability tests are listed below:
1. Generate VRP datasets at sizes:
* E.g., 50k, 100k, 500k, 1M.
2. Load each dataset into the RTR emulator.
3. For each dataset, measure:
* Full-table synchronization time.
* VRP update processing latency (from Section 6.1).
* ROV validation latency (from Section 6.2).
* System memory consumption.
* System CPU utilization during sync and steady state.
* Per-process utilization for the ROV process and the BGP or
routing process, when available.
4. Record failures
* Session drops
* Timeouts
* Missing VRPs
* ROV process crashes
Liu, et al. Expires 6 May 2027 [Page 17]
Internet-Draft ROV Benchmarking November 2026
5. Repeat 10 times per size for statistical stability.
6.5. VRP Churn and Stress Tests
*Objective*: Stress-test the DUT under rapid VRP changes to measure
stability, performance, and correctness.
The *test procedures* for VRP churn and stress tests are listed
below:
1. Baseline setup
* Load a stable VRP table (e.g., 500k).
* Establish the baseline BGP table with IPv4 and IPv6 route
counts stated explicitly.
2. Generate controlled churn patterns
* Rapid add or remove spikes: 100-10,000 VRPs per second.
* Sustained churn: continuous modifications for 5-10 minutes.
* Mixed churn: adds, removes, and changes simultaneously.
3. Measure DUT behavior
* VRP update backlog or queueing.
* ROV validation delays.
* System CPU spikes and, when available, spikes in the ROV
process and the BGP or routing process.
* BGP convergence degradation.
* Missed or dropped VRP updates.
4. Check correctness
* Verify that no stale or inconsistent ROV states remain.
5. Record crash, stall, or throttling events.
6.6. Resource Utilization
*Objective*: Measure resource consumption under various ROV
workloads.
Liu, et al. Expires 6 May 2027 [Page 18]
Internet-Draft ROV Benchmarking November 2026
The *test procedures* for RTR session behavior tests are listed
below:
1. Establish monitoring tools
* System CPU sampling (100-500 ms interval).
* System memory usage tracking.
* Per-process sampling for the ROV process and the BGP or
routing process, when available.
* Hardware counters if available.
2. Measure under conditions
* Idle ROV.
* Full VRP sync.
* VRP churn.
* BGP convergence triggered by ROV events.
3. Record
* System CPU load curves.
* Peak and steady-state system memory consumption.
* Process-level load curves for the ROV process and the BGP or
routing process, when available.
* Any evidence of saturation (e.g., 100% CPU, memory
exhaustion).
4. Identify thresholds
* Points where performance degrades or ROV processing becomes
unstable.
6.7. RTR Session Behavior Tests
*Objective*: Evaluate robustness and recovery of DUT under RTR
failure and failover scenarios, including both RTR session recovery
and the subsequent BGP recalculation caused by restored or changed
ROV state.
Liu, et al. Expires 6 May 2027 [Page 19]
Internet-Draft ROV Benchmarking November 2026
The *test procedures* for RTR session behavior tests are listed
below:
1. Session reset test
* Establish normal RTR session.
* Trigger forced session reset from emulator.
* Measure time to reestablish RTR session, ROV state rebuild
time, and time until validation state becomes consistent
again.
* After validation state consistency is restored, measure any
resulting BGP recalculation and convergence using the post-
installation BGP convergence timer defined in the BGP
convergence test.
2. Cache failover test
* Configure DUT with two RTR servers (primary + secondary).
* Terminate primary RTR connection.
* Measure failover time and data consistency after switch.
* Measure the time until affected BGP routes reach stable RIB
and FIB state after the restored or failover VRP state is
applied.
3. Full resynchronization timing
* From emulator, force a full Reset Query sequence.
* Measure full VRP reload time.
* Measure subsequent ROV-triggered BGP convergence when the
resynchronized VRP set changes validation states for installed
BGP routes.
* Compare across different VRP scales.
4. Incremental update performance
* Send controlled incremental PDUs.
* Measure processing latency and correctness.
Liu, et al. Expires 6 May 2027 [Page 20]
Internet-Draft ROV Benchmarking November 2026
* Introduce occasional malformed or unexpected PDUs to test
robustness.
7. Reporting Requirements
An ROV benchmarking report must provide enough details to allow
reproducibility and meaningful comparison across different DUTs.
Each report needs to include the following elements:
* *Test environment description*: The report must specify the DUT
hardware and software versions, the testbed topology, and all ROV-
related configuration parameters required to replicate the setup.
* *Input conditions*: The report must document the ROA count, VRP
count, VRP coverage relative to the announced routing space, VRP
overlap characteristics, the RIB-in size with IPv4 and IPv6 counts
stated separately, the presence and rate of VRP churn, and whether
RTR updates were incremental or full.
* *Metrics and results*: Each measured metric must include its
definition, a brief description of the measurement procedure, and
results presented in tabular numerical form (including minimum,
average, maximum, and at least P95 values). Graphs MAY be
included for clarification.
* *Deviations and anomalies*: Any deviation from the expected
behavior must be described, including the conditions under which
it occurred and whether the test was repeated.
* *Summary of observations*: The report must include a concise
summary of overall DUT performance, scalability limits observed,
and any significant effects of enabling ROV on BGP behavior.
In addition, the report is suggested to include, at minimum, the
following parameters:
* DUT hardware model, CPU architecture, memory size, and software
version.
* Complete DUT configuration relevant to ROV and BGP.
* Testbed topology description.
* Total number of ROAs and total number of resulting VRPs.
* VRP table size.
Liu, et al. Expires 6 May 2027 [Page 21]
Internet-Draft ROV Benchmarking November 2026
* VRP coverage relative to the announced routing space, including
the definition used for coverage.
* VRP overlap characteristics, including the probability or fraction
that a VRP is covered by another VRP.
* VRP churn rate.
* For VRP update tests, whether each update affects a single VRP or
multiple VRPs, and the operation type for each update.
* For validation-change tests, the validation-change type, such as
origin-AS mismatch, prefix-length violation, or loss/addition of a
covering VRP.
* For locality-sensitive tests, whether affected prefixes map to the
same or different regions of the validation structure, and the
locality rule used.
* RIB-in size, with IPv4 and IPv6 route counts stated separately.
* AFIs under test (IPv4, IPv6, or mixed).
* Number of RTR sessions.
* RTR timer configuration.
* Presence and parameters of data-plane traffic (if used), including
fixed frame size, PPS rate, traffic pattern, and address-selection
rule.
* ROV policy mode (e.g., reject Invalid).
* System CPU sampling interval and any process-level sampling
interval.
* Measurement repetition count.
* For BGP convergence tests, whether the reported timer is end-to-
end from VRP update transmission, post-installation BGP
convergence, or BGP-triggered convergence with ROV enabled/
disabled.
For each metric, the report must provide:
* Metric definition.
* Measurement method.
Liu, et al. Expires 6 May 2027 [Page 22]
Internet-Draft ROV Benchmarking November 2026
* Minimum, average, maximum, and at least P95 values.
* Number of samples collected.
8. Security Considerations
This document defines a benchmarking methodology for evaluating ROV
on routing devices. As such, it does not introduce new protocols,
modify existing security mechanisms, or create new vulnerabilities
within the RPKI system or BGP itself. All benchmarking activities
are intended to take place in isolated laboratory environments.
Nevertheless, a number of security considerations apply to the
execution and interpretation of the tests described in this document.
Benchmarking ROV necessarily involves the generation, manipulation,
and replay of RPKI objects. These test artifacts MUST NOT be
injected into production RPKI repositories, production RPKI caches,
or live BGP routing systems. Test-generated RPKI data sets SHOULD be
clearly separated from real-world trust anchors, and laboratory RPKI
caches SHOULD use isolated test Trust Anchors to prevent accidental
propagation.
Similarly, BGP routing information used in the tests including
simulated full tables, invalid prefixes, or artificially crafted
origin-AS combinations, MUST NOT leak into production routing
domains. All BGP sessions used for testing MUST be confined to a
closed environment without external connectivity.
Tests involving stress conditions, such as high churn rates or large-
scale VRP updates, may cause elevated CPU or memory consumption on
the DUT. Operators performing such tests SHOULD ensure that the DUT
is not simultaneously connected to any production network to avoid
unintended service degradation.
9. Implementation Status
This section summarizes the current implementation and experimental
coverage of this methodology, following the reporting approach in
[RFC7942]. It provides implementation experience for IETF
discussion.
A prototype implementation of the benchmarking methodology described
in this document supports the experiments reported in
Appendix "Appendix B. Implementation Details". It uses BIRD [BIRD]
as the routing and ROV implementation, the Linux kernel for
forwarding, a locally modified BRIO [BRIO] as the RTR source, and
Bngblaster [BNGBLASTER] for BGP traffic generation.
Liu, et al. Expires 6 May 2027 [Page 23]
Internet-Draft ROV Benchmarking November 2026
Implemented:
* Incremental VRP updates, full-table resynchronization, and
localized and distributed ROV validation tests.
* VRP-triggered BGP convergence and BGP-triggered convergence
comparisons with ROV enabled and disabled.
* VRP scalability, churn and stress, and resource-utilization tests.
* RTR session recovery, Cache Reset resynchronization, and
malformed-PDU handling tests.
Not yet implemented or incomplete:
* Comparison of localized and distributed ROV validation workloads
with the same number of BGP routes undergoing ROV state changes.
* Evaluation on additional routing implementations and dedicated
hardware routers; the reported experiments use a BIRD-based DUT
under WSL2.
* Primary-to-secondary RTR cache failover measurements. In the
current BIRD setup, both caches contribute authorizations
concurrently, so loss of one session does not exercise the
standby-cache activation required by this scenario.
* Hardware performance-counter measurements, which are unavailable
in the current WSL2 environment.
10. IANA Considerations
This document has no actions for IANA.
11. References
11.1. Normative References
[RFC2119] Bradner, S., "Key words for use in RFCs to Indicate
Requirement Levels", BCP 14, RFC 2119,
DOI 10.17487/RFC2119, March 1997,
<https://www.rfc-editor.org/rfc/rfc2119>.
[RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC
2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174,
May 2017, <https://www.rfc-editor.org/rfc/rfc8174>.
Liu, et al. Expires 6 May 2027 [Page 24]
Internet-Draft ROV Benchmarking November 2026
[RFC6811] Mohapatra, P., Scudder, J., Ward, D., Bush, R., and R.
Austein, "BGP Prefix Origin Validation", RFC 6811,
DOI 10.17487/RFC6811, January 2013,
<https://www.rfc-editor.org/rfc/rfc6811>.
[RFC8210] Bush, R. and R. Austein, "The Resource Public Key
Infrastructure (RPKI) to Router Protocol, Version 1",
RFC 8210, DOI 10.17487/RFC8210, September 2017,
<https://www.rfc-editor.org/rfc/rfc8210>.
[RFC1242] Bradner, S., "Benchmarking Terminology for Network
Interconnection Devices", RFC 1242, DOI 10.17487/RFC1242,
July 1991, <https://www.rfc-editor.org/rfc/rfc1242>.
[RFC2285] Mandeville, R., "Benchmarking Terminology for LAN
Switching Devices", RFC 2285, DOI 10.17487/RFC2285,
February 1998, <https://www.rfc-editor.org/rfc/rfc2285>.
[RFC2544] Bradner, S. and J. McQuaid, "Benchmarking Methodology for
Network Interconnect Devices", RFC 2544,
DOI 10.17487/RFC2544, March 1999,
<https://www.rfc-editor.org/rfc/rfc2544>.
[RFC2889] Mandeville, R. and J. Perser, "Benchmarking Methodology
for LAN Switching Devices", RFC 2889,
DOI 10.17487/RFC2889, August 2000,
<https://www.rfc-editor.org/rfc/rfc2889>.
[RFC3918] Stopp, D. and B. Hickman, "Methodology for IP Multicast
Benchmarking", RFC 3918, DOI 10.17487/RFC3918, October
2004, <https://www.rfc-editor.org/rfc/rfc3918>.
11.2. Informative References
[RFC7942] Sheffer, Y. and A. Farrel, "Improving Awareness of Running
Code: The Implementation Status Section", BCP 205,
RFC 7942, DOI 10.17487/RFC7942, July 2016,
<https://www.rfc-editor.org/rfc/rfc7942>.
[BIRD] CZ.NIC, "BIRD Internet Routing Daemon",
<https://bird.nic.cz/>.
[BRIO] National Institute of Standards and Technology, "NIST BGP-
RPKI-IO (BRIO)", <https://github.com/usnistgov/NIST-BRIO>.
[BNGBLASTER]
RtBrick, "Routing Protocol and BNG Blaster",
<https://rtbrick.github.io/bngblaster/>.
Liu, et al. Expires 6 May 2027 [Page 25]
Internet-Draft ROV Benchmarking November 2026
[RIPE-RIS] RIPE NCC, "Routing Information Service (RIS)",
<https://www.ripe.net/analyse/internet-measurements/
routing-information-service-ris/>.
[RPKIViews]
Snijders, J., "RPKIViews", <https://rpkiviews.org/>.
[ROUTINATOR]
NLnet Labs, "Routinator",
<https://github.com/NLnetLabs/routinator>.
[STAYRTR] "StayRTR", <https://github.com/bgp/stayrtr>.
Acknowledgements
Many thanks to Giuseppe Fioccola, Christian Giese, and Maria Matejka
for their review, comments, and suggestions.
Appendix A. Summary of Changes
This appendix summarizes the major changes introduced during the
evolution of this document.
Versions -00 through -03 were published under the name draft-lg-bmwg-
benchmarking-methodology-for-rov. Following the consolidation of the
BMWG and IPPM working groups into the Benchmarking and Performance
Measurement (BPM) Working Group, the document was renamed to draft-
lg-bpm-benchmarking-methodology-for-rov. Accordingly, draft-lg-bpm-
benchmarking-methodology-for-rov-00 continues the work of draft-lg-
bmwg-benchmarking-methodology-for-rov-03.
Appendix A.1. Changes from draft-lg-bmwg-benchmarking-methodology-for-
rov-03 to draft-lg-bpm-benchmarking-methodology-for-rov-00
The major changes introduced in draft-lg-bpm-benchmarking-
methodology-for-rov-00, compared with draft-lg-bmwg-benchmarking-
methodology-for-rov-03, are as follows:
* Added Appendix B documenting a BIRD-based implementation of the
benchmarking methodology and its experimental results.
* Added an Implementation Status section describing the prototype
and its experimental coverage.
Liu, et al. Expires 6 May 2027 [Page 26]
Internet-Draft ROV Benchmarking November 2026
Appendix A.2. Changes from -02 to -03
Version -03 adds this appendix to summarize the major changes
introduced across document revisions and improve revision
traceability. No technical changes to the benchmarking methodology
or benchmark procedures are introduced in this revision.
Appendix A.3. Changes from -01 to -02
The major changes from version -01 to version -02 are as follows:
* Expanded the controlled input conditions to distinguish ROA and
VRP populations and to characterize VRP coverage relative to the
announced routing space and overlap properties within the VRP
dataset.
* Refined the ROV update processing latency test to distinguish
single-VRP and batch updates, document addition, withdrawal, and
modification operations and the expected validation-state
transitions, and report first-VRP installation latency, last-VRP
installation latency, and aggregate batch processing time when
observable.
* Refined the ROV validation latency test by introducing the
structural distribution of affected prefixes as a controlled
variable. The methodology now distinguishes localized and
distributed prefix sets and requires the locality rule, AFI mix,
prefix-length distribution, and cause of each validation-state
change to be documented.
* Reworked the BGP convergence benchmark into two cases: VRP-
triggered BGP convergence and BGP-triggered convergence with and
without ROV. The revision also introduced separate end-to-end and
post-VRP-installation convergence timers and more explicit
controls for comparing ROV-enabled and ROV-disabled runs.
* Extended RTR session recovery, cache failover, and full
resynchronization tests to include the subsequent BGP
recalculation and convergence caused by restored or changed ROV
state.
* Expanded the reporting requirements to cover the newly introduced
VRP dataset characteristics, update types, validation-change
causes, locality conditions, and convergence timer definitions,
and refined the use of normative requirement language throughout
the methodology.
Liu, et al. Expires 6 May 2027 [Page 27]
Internet-Draft ROV Benchmarking November 2026
Appendix A.4. Changes from -00 to -01
The major changes from version -00 to version -01 are as follows:
* Refined the test topology and baseline BGP input conditions to
distinguish IPv4 and IPv6 route counts and to support mixed or
AFI-specific benchmarking under fixed and repeatable conditions.
* Refined the data-plane traffic profile for convergence
measurements by specifying a fixed frame size, PPS-based constant-
rate traffic, separate treatment of burst traffic, and
deterministic source and destination address selection rules.
* Clarified resource utilization metrics and measurement procedures
by distinguishing system-level CPU and memory consumption from
per-process utilization of the ROV process and the BGP or routing
process, when such counters are available.
* Updated the scalability, churn, resource utilization, and
reporting procedures to consistently document AFIs under test,
data-plane traffic parameters, and system-level and process-level
resource measurements.
Appendix B. Implementation Details
This appendix describes the application of the benchmark tests in
Section 6 to a BIRD-based DUT. It presents the test environment,
shared input datasets, measurement procedures, and results.
Appendix B.1. Test Environment
The DUT consists of BIRD for routing and ROV processing, together
with the Linux kernel for packet forwarding. BRIO supplies VRPs over
the RTR protocol, with local modifications providing first-Prefix-PDU
send timestamps and controlled client disconnection. Bngblaster
establishes BGP sessions and advertises the baseline and update
routes, with a local modification for IPv6 neighbor emulation. A
separate ingress sends UDP probes through the DUT; receivers at the
emulated BGP peers identify the forwarding next hop. The test
environment configuration is summarized in Figure 2.
The test setup uses BRIO for direct control over incremental VRP
additions and withdrawals in the VRP churn and stress test
(Section 6.5). By comparison, widely used RTR tools, such as
Routinator [ROUTINATOR] and StayRTR [STAYRTR], derive incremental
updates by comparing complete VRP sets internally. The additional
processing delays transmission, making this approach unsuitable for
continuous monitoring in this test.
Liu, et al. Expires 6 May 2027 [Page 28]
Internet-Draft ROV Benchmarking November 2026
+==========================+=================================================+
| Component | Configuration |
+==========================+=================================================+
| Operating environment | Linux under WSL2; kernel |
| | `6.18.33.2-microsoft-standard-WSL2` |
+--------------------------+-------------------------------------------------+
| CPU | Intel Core 5 220H |
+--------------------------+-------------------------------------------------+
| Memory available to WSL2 | 15.37 GiB |
+--------------------------+-------------------------------------------------+
| Routing software | BIRD 2.19.2 |
+--------------------------+-------------------------------------------------+
| BGP traffic generator | Bngblaster 0.9.39, with local modifications |
| | |
+--------------------------+-------------------------------------------------+
| RTR source | BRIO b7.1.2, with local modifications |
+--------------------------+-------------------------------------------------+
Figure 2: Test environment configuration.
Appendix B.2. BGP and VRP Data
*Initial BGP data.* Routes are sampled from the RIPE RIS [RIPE-RIS]
RRC00 MRT snapshot of 9 September 2026 at 00:00 UTC. Sampling is
performed separately for IPv4 and IPv6, with quotas allocated across
peers, targeting 1,000,000 IPv4 and 250,000 IPv6 routes. Conversion
preserves the original BGP path attributes. Two filters are applied:
* Routes whose NEXT_HOP differs from the advertising peer's address
are excluded to avoid exceeding the interface capacity of the test
setup when emulating additional next hops. The Bngblaster build
supports at most 255 TCP-enabled interfaces.
* Routes with AS_PATH text too long to be displayed in full are
excluded to prevent truncation in BIRD's CLI output and RIB change
logs.
The resulting baseline contains 999,125 IPv4 and 222,006 IPv6
candidate routes from 99 peers, covering 805,052 distinct prefixes.
Different peers' advertisements of the same prefix are retained as
separate candidates. All experiments use this BGP baseline.
Liu, et al. Expires 6 May 2027 [Page 29]
Internet-Draft ROV Benchmarking November 2026
*Initial VRP data.* VRPs originate from the RPKIViews [RPKIViews]
snapshot of 8 September 2026 at 23:47:23 UTC. AS0 records are
removed, and additional authorizations are generated to bring the
numbers of Invalid and NotFound BGP routes closer together. The
resulting baseline contains 1,060,368 VRPs: 820,699 IPv4 and 239,669
IPv6. Experiments use either this complete baseline or sampled
subsets.
The inputs comprise four datasets:
* *Single-batch VRP updates:* Nested initial sets of 50k, 100k,
500k, and 1M VRPs. Randomly select BGP candidates and change
their ROV states until reaching 1,000 VRP operations.
* *Multi-batch VRP updates:* Apply the same generation method over
300 batches, starting with 500k VRPs and targeting 100, 500, or
1,000 operations per batch.
* *Prefix-selected VRP updates:* Select 100, 500, or 1,000 distinct
prefixes and change a selected candidate's ROV state for each.
* *BGP updates:* Keep the complete VRP baseline fixed and generate
1,000 BGP operations per workload. Each operation is selected to
change the best path in at least one of the configurations with
ROV enabled or disabled.
A single ROV transition may require multiple VRP operations and
affect additional routes. Consequently, a prefix-selected workload
targeting 1,000 prefixes may require more than 1,000 VRP operations,
and the resulting ROV state changes may extend to BGP routes for
other prefixes.
The ROV validation latency test in Section 6.2 requires both
localized and distributed scenarios. The localized scenario draws
prefixes from as few IPv4 /8 or IPv6 /32 address regions as needed to
meet the target count. The distributed scenario samples prefixes
randomly from the entire eligible pool without restricting address
regions.
The BGP update dataset includes mixed workloads and workloads
containing a single operation type. The five operation types are:
* *Withdraw best:* Withdraw the existing best route.
* *Add Valid:* Announce a new route with a Valid ROV state that can
become the best path.
Liu, et al. Expires 6 May 2027 [Page 30]
Internet-Draft ROV Benchmarking November 2026
* *Add Invalid:* Announce a new Invalid route that can become the
best path.
* *Invalid to Valid:* Change an existing Invalid route to Valid so
that it becomes the best path.
* *Best to Invalid:* Change the existing best route to Invalid.
Appendix B.3. Measurement Methods
The collection methods and observation timestamps for authorization
installation, validation queries, routing decisions, and forwarding
behavior are summarized in Figure 3.
+========================+=========================+=========================+
| Observable | Collection method | Observation time |
+========================+=========================+=========================+
| VRP installation or | BIRD VRP change logs | Timestamp associated |
| withdrawal | | with the observed log |
| | | event |
+------------------------+-------------------------+-------------------------+
| ROV state | BIRD CLI queries for a | Receipt of the query |
| | prefix and origin ASN | result |
+------------------------+-------------------------+-------------------------+
| RIB changes | BIRD best-path change | Timestamp associated |
| | logs | with the observed log |
| | | event |
+------------------------+-------------------------+-------------------------+
| FIB changes | UDP probes sent through | Receipt of a probe, or |
| | a separate ingress and | expiry of its |
| | received at candidate | observation deadline |
| | next hops | where applicable |
+------------------------+-------------------------+-------------------------+
Figure 3: Measurement methods and observation timestamps.
The collector queries ROV states through BIRD's command-line
interface over a persistent UNIX domain socket connection. It sends
query commands in groups and receives all replies before submitting
the next group. In the ordinary convergence experiments, a new
observation is compared with the previous state, and only a state
change creates an event.
Liu, et al. Expires 6 May 2027 [Page 31]
Internet-Draft ROV Benchmarking November 2026
FIB probes are sent continuously at the configured aggregate packet
rate, without waiting for the preceding probe to arrive. In the
ordinary convergence experiments, a probe with no reception within
one second produces a no-reception observation. The observed
forwarding state is compared with the previous state, and only a
change creates a new event.
Some experiments measure short intervals that are particularly
sensitive to pauses in the measurement process. The measurement
software therefore disables Python's automatic garbage collection
during timed collection in all experiments and restores its previous
setting afterward.
Appendix B.4. General Conventions
ROV queries use a default group size of eight commands. Convergence
experiments use a ten-second quiet period and, where forwarding is
monitored, an aggregate probe limit of 10,000 packets/s with a one-
second packet deadline. Collection stops after the quiet period,
with pending logs and events drained. In ordinary convergence
experiments, expected-state matching and final correctness checks are
performed after timed collection. The reported last-change endpoint
excludes the subsequent quiet period. The VRP churn and stress test
(Section 6.5) retains pending targets across subsequent batches and
checks their completion periodically during collection. The RTR
session behavior tests (Section 6.7) additionally require scenario-
specific protocol events.
The default origin for transmission-based intervals is BRIO's
monotonic timestamp immediately before the first successful Prefix
PDU send call for the relevant response. Continuous-update VRP
latency measurements in Figure 14 instead start at the corresponding
Serial Notify send time: multiple batches may share one RTR response,
preventing its first Prefix PDU from providing a separate origin for
each batch. ROV, RIB, and FIB latencies in that experiment use a
post-installation origin. Post-installation intervals and scenario-
specific origins are defined in the corresponding experiments.
Latency values are expressed in milliseconds unless another unit is
specified. Standard deviation is the sample standard deviation. P95
is the 95th percentile using linear interpolation between ordered
samples. The sample count denotes repeated runs, different inputs,
or observed batches as specified in each section. An em dash denotes
an unavailable measurement; N/A denotes an inapplicable measurement.
Liu, et al. Expires 6 May 2027 [Page 32]
Internet-Draft ROV Benchmarking November 2026
Appendix B.5. ROV Update Processing Latency
This experiment implements the update-processing test in Section 6.1
using nested initial sets of 50k, 100k, 500k, and 1M VRPs, with 1,000
VRP operations at each scale. It measures installation latency for
incremental and full-table updates. Each condition is repeated ten
times. The results are shown in Figure 4.
+============+===================+=================+===================+===================+===================+
| Initial | Mean (ms) | Std. dev. (ms) | Minimum (ms) | Maximum (ms) | P95 (ms) |
| VRPs | | | | | |
+------------+--------+----------+--------+--------+--------+----------+--------+----------+--------+----------+
| | Incr. | Full | Incr. | Full | Incr. | Full | Incr. | Full | Incr. | Full |
+============+========+==========+========+========+========+==========+========+==========+========+==========+
| 50k | 47.489 | 160.076 | 2.434 | 3.513 | 45.186 | 154.845 | 51.302 | 164.765 | 51.170 | 164.759 |
+------------+--------+----------+--------+--------+--------+----------+--------+----------+--------+----------+
| 100k | 46.460 | 319.486 | 3.047 | 7.385 | 43.172 | 305.609 | 51.389 | 331.359 | 51.323 | 329.472 |
+------------+--------+----------+--------+--------+--------+----------+--------+----------+--------+----------+
| 500k | 42.628 | 1631.664 | 2.713 | 16.436 | 38.670 | 1598.010 | 47.795 | 1659.967 | 47.325 | 1653.157 |
+------------+--------+----------+--------+--------+--------+----------+--------+----------+--------+----------+
| 1M | 38.704 | 3366.851 | 3.245 | 54.778 | 32.703 | 3299.205 | 44.141 | 3447.183 | 43.178 | 3444.970 |
+------------+--------+----------+--------+--------+--------+----------+--------+----------+--------+----------+
Figure 4: VRP installation latency for incremental and full-table
updates.
Incr. denotes incremental delivery; Full denotes full-table
resynchronization. In Figure 4, incremental update latency is short
and similar across the tested table sizes, averaging approximately
40–50 ms. Full-table resynchronization latency increases as the VRP
table grows.
Appendix B.6. ROV Validation Latency
This experiment implements the validation-latency test in Section 6.2
using 1,060,368 initial VRPs and updates targeting 100, 500, or 1,000
selected prefixes in localized and distributed groups. With BIRD's
ROA-table trie enabled, it measures validation latency from VRP
installation completion. The results are shown in Figure 5.
Liu, et al. Expires 6 May 2027 [Page 33]
Internet-Draft ROV Benchmarking November 2026
+============+=====================+=====================+=====================+=====================+=====================+
| Selected | Mean (ms) | Std. dev. (ms) | Minimum (ms) | Maximum (ms) | P95 (ms) |
| prefixes | | | | | |
+------------+----------+----------+----------+----------+----------+----------+----------+----------+----------+----------+
| | Local. | Distrib. | Local. | Distrib. | Local. | Distrib. | Local. | Distrib. | Local. | Distrib. |
+============+==========+==========+==========+==========+==========+==========+==========+==========+==========+==========+
| 100 | 71.847 | 104.377 | 5.244 | 17.235 | 66.315 | 87.147 | 79.560 | 141.479 | 78.635 | 134.556 |
+------------+----------+----------+----------+----------+----------+----------+----------+----------+----------+----------+
| 500 | 133.143 | 622.376 | 11.623 | 38.711 | 118.246 | 565.663 | 157.540 | 705.357 | 151.338 | 681.772 |
+------------+----------+----------+----------+----------+----------+----------+----------+----------+----------+----------+
| 1,000 | 174.169 | 2119.743 | 33.637 | 961.166 | 144.346 | 906.791 | 256.348 | 3934.232 | 231.466 | 3501.198 |
+------------+----------+----------+----------+----------+----------+----------+----------+----------+----------+----------+
Figure 5: ROV validation latency for localized and distributed
workloads.
Local. denotes localized inputs; Distrib. denotes distributed inputs.
As shown in Figure 5, ROV validation latency increases with the
number of selected prefixes in both scenarios. For the same selected
prefix count, localized workloads show lower observed latency.
However, selecting the same number of prefixes does not mean that the
same number of prefixes is actually affected. For example, both
workloads select 1,000 prefixes, but ROV states change for 3,640
prefixes in the localized scenario and 23,777 prefixes in the
distributed scenario. The longer latency in the distributed scenario
is therefore associated with both a different prefix distribution and
more prefixes to process.
Appendix B.7. BGP Convergence with ROV Enabled
The following experiments correspond to the two cases in Section 6.3.
Appendix B.7.1. Case A: VRP-triggered BGP convergence
This experiment implements the test in Section 6.3.1 using 1,060,368
initial VRPs and the distributed workload with 1,000 selected
prefixes, requiring 1,230 VRP operations. It monitors RIB and FIB
changes. End-to-end and post-installation convergence times are
shown in Figure 6.
Liu, et al. Expires 6 May 2027 [Page 34]
Internet-Draft ROV Benchmarking November 2026
+===================+==========+=========+==========+==========+==========+
| Timer | Mean | Std. | Minimum | Maximum | P95 (ms) |
| | (ms) | dev. | (ms) | (ms) | |
| | | (ms) | | | |
+===================+==========+=========+==========+==========+==========+
| End-to-end | 5027.552 | 267.088 | 4826.382 | 5736.282 | 5469.557 |
| convergence | | | | | |
+-------------------+----------+---------+----------+----------+----------+
| Post-installation | 4813.556 | 363.372 | 4540.745 | 5703.936 | 5432.527 |
| BGP convergence | | | | | |
+-------------------+----------+---------+----------+----------+----------+
Figure 6: VRP-triggered BGP convergence.
The final FIB change determines the convergence endpoint in all ten
runs. Mean end-to-end convergence in Figure 6 is 5.03 s, of which
4.81 s occurs after VRP installation. Thus, most of the measured
convergence interval follows installation, with FIB convergence
taking longer than RIB convergence.
Appendix B.7.2. Case B: BGP-triggered convergence with and without ROV
This experiment implements the test in Section 6.3.2 with a fixed set
of 1,060,368 VRPs and 1,000 BGP operations per workload. One mixed
workload and five individual operation-type workloads are replayed
with ROV enabled and disabled. RIB and FIB convergence is measured
from the first BGP UPDATE send to the last observed change. The
results are shown in Figure 7 through Figure 12.
*Mixed.*
+============+===========+==============+==============+===========+
| ROV policy | Mean (ms) | Minimum (ms) | Maximum (ms) | P95 (ms) |
+============+===========+==============+==============+===========+
| Enabled | 10591.995 | 10532.102 | 10664.151 | 10662.097 |
+------------+-----------+--------------+--------------+-----------+
| Disabled | 10580.443 | 10539.504 | 10648.550 | 10632.916 |
+------------+-----------+--------------+--------------+-----------+
Figure 7: BGP convergence with ROV enabled and disabled: mixed
updates.
*Withdraw best.*
Liu, et al. Expires 6 May 2027 [Page 35]
Internet-Draft ROV Benchmarking November 2026
+============+===========+==============+==============+===========+
| ROV policy | Mean (ms) | Minimum (ms) | Maximum (ms) | P95 (ms) |
+============+===========+==============+==============+===========+
| Enabled | 9791.912 | 9168.918 | 11690.974 | 11275.438 |
+------------+-----------+--------------+--------------+-----------+
| Disabled | 9599.031 | 9066.639 | 10978.189 | 10714.374 |
+------------+-----------+--------------+--------------+-----------+
Figure 8: BGP convergence with ROV enabled and disabled: Withdraw
best.
*Add Valid.*
+============+===========+==============+==============+===========+
| ROV policy | Mean (ms) | Minimum (ms) | Maximum (ms) | P95 (ms) |
+============+===========+==============+==============+===========+
| Enabled | 10158.053 | 9685.791 | 10586.965 | 10506.822 |
+------------+-----------+--------------+--------------+-----------+
| Disabled | 10361.078 | 10173.760 | 11069.728 | 10893.730 |
+------------+-----------+--------------+--------------+-----------+
Figure 9: BGP convergence with ROV enabled and disabled: Add Valid.
*Add Invalid.*
+============+===========+==============+==============+===========+
| ROV policy | Mean (ms) | Minimum (ms) | Maximum (ms) | P95 (ms) |
+============+===========+==============+==============+===========+
| Enabled | N/A | N/A | N/A | N/A |
+------------+-----------+--------------+--------------+-----------+
| Disabled | 9962.983 | 9685.891 | 10177.751 | 10176.955 |
+------------+-----------+--------------+--------------+-----------+
Figure 10: BGP convergence with ROV enabled and disabled: Add
Invalid.
*Invalid to Valid.*
+============+===========+==============+==============+==========+
| ROV policy | Mean (ms) | Minimum (ms) | Maximum (ms) | P95 (ms) |
+============+===========+==============+==============+==========+
| Enabled | 8686.659 | 8111.456 | 9856.441 | 9621.860 |
+------------+-----------+--------------+--------------+----------+
| Disabled | 8011.848 | 8005.596 | 8031.975 | 8027.183 |
+------------+-----------+--------------+--------------+----------+
Figure 11: BGP convergence with ROV enabled and disabled: Invalid
to Valid.
Liu, et al. Expires 6 May 2027 [Page 36]
Internet-Draft ROV Benchmarking November 2026
*Best to Invalid.*
+============+===========+==============+==============+==========+
| ROV policy | Mean (ms) | Minimum (ms) | Maximum (ms) | P95 (ms) |
+============+===========+==============+==============+==========+
| Enabled | 9298.554 | 9105.652 | 9568.952 | 9558.411 |
+------------+-----------+--------------+--------------+----------+
| Disabled | 8385.576 | 8006.559 | 9898.653 | 9520.511 |
+------------+-----------+--------------+--------------+----------+
Figure 12: BGP convergence with ROV enabled and disabled: Best to
Invalid.
The mixed workload produces similar convergence times with ROV
enabled and disabled (Figure 7). For Withdraw best and Add Valid,
the operations have similar effects on route selection in both
configurations, and the measured convergence times are also similar
(Figure 8 and Figure 9). For Add Invalid, ROV rejects the added
routes, so no RIB or forwarding changes are observed with ROV enabled
and no last-change latency is reported (Figure 10). The larger
differences occur for Invalid to Valid (Figure 11) and Best to
Invalid (Figure 12): the validation-state transitions do not change
route eligibility with ROV disabled, whereas they require route
reselection with ROV enabled. These two workloads therefore show
larger differences in convergence time between the configurations.
Appendix B.8. VRP Scalability Tests
This experiment implements the scalability test in Section 6.4 using
50k, 100k, 500k, and 1M initial VRPs, with 1,000 VRP operations at
each scale. It measures initial synchronization, VRP update latency,
validation latency after VRP installation, and BIRD CPU and memory
consumption. Each condition is repeated ten times. The results are
shown in Figure 13.
Liu, et al. Expires 6 May 2027 [Page 37]
Internet-Draft ROV Benchmarking November 2026
+==========+=================+==========+============+===========+===========+
| Initial | Full-table | VRP | ROV | BIRD CPU | BIRD RSS |
| VRPs | synchronization | update | validation | (%) | (MiB) |
| | (ms) | latency | latency | | |
| | | (ms) | (ms) | | |
+==========+=================+==========+============+===========+===========+
| 50k | 17.104 | 48.578 | 206.269 | 38.366 | 462.971 |
+----------+-----------------+----------+------------+-----------+-----------+
| 100k | 32.779 | 49.258 | 340.882 | 58.298 | 466.205 |
+----------+-----------------+----------+------------+-----------+-----------+
| 500k | 188.375 | 44.607 | 789.480 | 66.257 | 501.640 |
+----------+-----------------+----------+------------+-----------+-----------+
| 1M | 422.649 | 41.032 | 3099.493 | 70.451 | 562.093 |
+----------+-----------------+----------+------------+-----------+-----------+
Figure 13: VRP scalability measurements.
The CPU column in Figure 13 reports BIRD utilization during initial
synchronization, with one fully occupied logical CPU equal to 100%.
Memory is BIRD's steady-state resident set size (RSS). Values are
reported as means across runs. In this test setup, ROV and BGP
processing share a single BIRD process, so their resource usage
cannot be measured separately and is reported for BIRD as a whole.
As shown in Figure 13, initial synchronization and post-installation
ROV validation take longer as the VRP table grows. Mean incremental
VRP update latency ranges from 41.032 to 49.258 ms, similar to the
update-processing results in Figure 4. Mean BIRD CPU utilization
during synchronization rises from 38.4% to 70.5%, while steady-state
RSS increases from 463 MiB to 562 MiB. RSS varies little across
repeated runs at each scale.
Appendix B.9. VRP Churn and Stress Tests
This experiment implements the churn test in Section 6.5 with 500k
initial VRPs and 300 batches at one-second intervals, targeting 100,
500, or 1,000 VRP operations per batch. It measures VRP, ROV, RIB,
and FIB latency under continuous updates. Only targets predicted to
change once over the complete update sequence are monitored.
Delivery delays and aggregation of multiple batches into one RTR
response can combine or cancel repeated operations. Within the DUT,
subsequent updates may also arrive before earlier changes have
propagated to routing and forwarding, so intermediate states may
produce neither a RIB change log nor a change in the forwarding next
hop observed by probes. Targets with repeated changes are therefore
excluded from measurement; their updates remain in the submitted
workload.
Liu, et al. Expires 6 May 2027 [Page 38]
Internet-Draft ROV Benchmarking November 2026
In practice, continuous updates substantially affected the
measurement process. Large numbers of ROV and FIB targets increased
the query and probe workload, while high load on the DUT delayed CLI
responses and could also delay packet forwarding. BIRD's ROA-table
settle timers also defer route revalidation after VRP updates, adding
a waiting period before routing and forwarding changes. The
measurement procedure therefore includes the following measures to
reduce these effects:
* Monitor selected batches rather than every batch: ROV, RIB, and
FIB targets are selected from batches 10, 20, and every subsequent
tenth batch through batch 300. VRP logs are collected for all
batches.
* Measure ROV and FIB in separate runs to reduce interference
between CLI queries and forwarding probes.
* Set ROA-table settle timers to zero for RIB and FIB measurements
to remove the additional waiting period before route revalidation.
* Check completion every five seconds and remove completed ROV/FIB
targets to avoid continuing queries and probes for those targets.
Monitoring continues across batch boundaries until the expected state
is observed. For each component, the batch endpoint is the latest of
the selected targets' first observations of their expected resulting
states. A complete batch latency requires all selected targets to be
observed. Collection ends after all updates have been submitted and
all selected targets have completed, or when the overall timeout
expires; final correctness checks follow collection.
VRP latency starts at the Serial Notify send time because multiple
batches may be combined into one RTR response, whose first Prefix PDU
cannot provide a separate timing origin for each batch. ROV, RIB,
and FIB post-installation latencies start at the last observed VRP
change for the corresponding batch. All 30 selected batches at each
size have complete ROV, RIB, and FIB matches.
VRP latency is shown in Figure 14.
Liu, et al. Expires 6 May 2027 [Page 39]
Internet-Draft ROV Benchmarking November 2026
+============+=============+=============+=============+=============+=============+
| Target | Mean | Std. dev. | Minimum | Maximum | P95 |
| operations | (ms) | (ms) | (ms) | (ms) | (ms) |
| per batch | | | | | |
+============+=============+=============+=============+=============+=============+
| 100 | 69.234 | 121.791 | 41.754 | 1856.278 | 61.636 |
+------------+-------------+-------------+-------------+-------------+-------------+
| 500 | 74.930 | 172.700 | 42.722 | 2768.577 | 66.831 |
+------------+-------------+-------------+-------------+-------------+-------------+
| 1,000 | 78.477 | 138.877 | 42.920 | 1901.677 | 76.311 |
+------------+-------------+-------------+-------------+-------------+-------------+
Figure 14: VRP latency under continuous updates.
Post-installation ROV latency is shown in Figure 15.
+============+=============+=============+=============+=============+=============+
| Target | Mean | Std. dev. | Minimum | Maximum | P95 |
| operations | (ms) | (ms) | (ms) | (ms) | (ms) |
| per batch | | | | | |
+============+=============+=============+=============+=============+=============+
| 100 | 348.373 | 333.814 | 10.100 | 1736.384 | 898.628 |
+------------+-------------+-------------+-------------+-------------+-------------+
| 500 | 755.901 | 410.524 | 309.312 | 2469.823 | 1403.662 |
+------------+-------------+-------------+-------------+-------------+-------------+
| 1,000 | 1477.077 | 869.270 | 227.354 | 3544.283 | 3031.084 |
+------------+-------------+-------------+-------------+-------------+-------------+
Figure 15: Post-installation ROV latency under continuous updates.
Post-installation RIB latency is shown in Figure 16.
+============+=============+=============+=============+=============+=============+
| Target | Mean | Std. dev. | Minimum | Maximum | P95 |
| operations | (ms) | (ms) | (ms) | (ms) | (ms) |
| per batch | | | | | |
+============+=============+=============+=============+=============+=============+
| 100 | 2497.838 | 249.244 | 1683.396 | 3129.184 | 2838.429 |
+------------+-------------+-------------+-------------+-------------+-------------+
| 500 | 3538.866 | 578.829 | 2398.800 | 6240.222 | 3848.731 |
+------------+-------------+-------------+-------------+-------------+-------------+
| 1,000 | 3412.681 | 518.784 | 2099.247 | 4159.981 | 4117.594 |
+------------+-------------+-------------+-------------+-------------+-------------+
Figure 16: Post-installation RIB latency under continuous updates.
Post-installation FIB latency is shown in Figure 17.
Liu, et al. Expires 6 May 2027 [Page 40]
Internet-Draft ROV Benchmarking November 2026
+============+=============+=============+=============+=============+=============+
| Target | Mean | Std. dev. | Minimum | Maximum | P95 |
| operations | (ms) | (ms) | (ms) | (ms) | (ms) |
| per batch | | | | | |
+============+=============+=============+=============+=============+=============+
| 100 | 3464.875 | 259.310 | 2677.374 | 4132.012 | 3822.033 |
+------------+-------------+-------------+-------------+-------------+-------------+
| 500 | 4552.836 | 620.005 | 3435.884 | 7504.002 | 4888.761 |
+------------+-------------+-------------+-------------+-------------+-------------+
| 1,000 | 4638.344 | 510.229 | 3190.433 | 5435.035 | 5357.773 |
+------------+-------------+-------------+-------------+-------------+-------------+
Figure 17: Post-installation FIB latency under continuous updates.
Additional single-batch tests measure VRP installation, ROV
validation, and BGP convergence latency. Each test initializes the
corresponding pre-update state, applies one batch, and measures VRP,
ROV, RIB, and FIB latency. ROV and FIB are measured in separate
runs, consistent with the continuous-update tests. The latency
comparison is shown in Figure 18.
+============+===================+===================+===================+===================+
| Operations | VRP latency | Post-installation | Post-installation | Post-installation |
| per batch | | ROV latency | RIB latency | FIB latency |
| | Mean (ms) | Mean (ms) | Mean (ms) | Mean (ms) |
+------------+---------+---------+---------+---------+---------+---------+---------+---------+
| | Cont. | Single | Cont. | Single | Cont. | Single | Cont. | Single |
+============+=========+=========+=========+=========+=========+=========+=========+=========+
| 100 | 69.234 | 44.654 | 348.373 | 20.659 | 2497.838| 1839.835| 3464.875| 2834.748|
+------------+---------+---------+---------+---------+---------+---------+---------+---------+
| 500 | 74.930 | 45.496 | 755.901 | 48.868 | 3538.866| 2202.871| 4552.836| 3254.135|
+------------+---------+---------+---------+---------+---------+---------+---------+---------+
| 1,000 | 78.477 | 46.229 | 1477.077| 63.922 | 3412.681| 2110.765| 4638.344| 3174.544|
+------------+---------+---------+---------+---------+---------+---------+---------+---------+
Figure 18: Latency comparison between continuous and single-batch
updates.
Figure 18 shows higher VRP, ROV, RIB, and FIB latencies under
continuous updates than under single-batch updates. However, the
measurement process itself also experienced delays during continuous
updates. A monitoring round queries or probes all current targets
once. Figure 19 compares the average duration and target count of
these rounds between continuous and single-batch tests. The averages
use the first complete round after each selected batch adds its
targets, including any unfinished targets from earlier batches.
Liu, et al. Expires 6 May 2027 [Page 41]
Internet-Draft ROV Benchmarking November 2026
+============+==============+=============+==============+=============+==============+
| Operations | Update mode | ROV round | ROV targets | FIB round | FIB targets |
| per batch | | time (ms) | per round | time (ms) | per round |
+============+==============+=============+==============+=============+==============+
| 100 | Continuous | 269.888 | 321.933 | 37.509 | 162.300 |
+------------+--------------+-------------+--------------+-------------+--------------+
| 100 | Single-batch | 24.745 | 321.933 | 31.968 | 162.300 |
+------------+--------------+-------------+--------------+-------------+--------------+
| 500 | Continuous | 729.705 | 608.267 | 90.170 | 370.533 |
+------------+--------------+-------------+--------------+-------------+--------------+
| 500 | Single-batch | 46.149 | 608.267 | 71.983 | 367.367 |
+------------+--------------+-------------+--------------+-------------+--------------+
| 1,000 | Continuous | 2502.206 | 1827.533 | 370.557 | 1046.067 |
+------------+--------------+-------------+--------------+-------------+--------------+
| 1,000 | Single-batch | 60.419 | 741.667 | 93.249 | 494.233 |
+------------+--------------+-------------+--------------+-------------+--------------+
Figure 19: Monitoring round duration and target count for
continuous and single- batch updates.
As shown in Figure 19, at 100 and 500 operations per batch, each ROV
query round covers the same number of targets in continuous and
single-batch tests, but takes substantially longer in the continuous
tests. At 1,000 operations per batch, unfinished targets from
earlier batches remain under monitoring alongside new targets; both
ROV and FIB rounds cover more targets and take longer than in the
single-batch tests. Together, Figure 18 and Figure 19 show higher
observed latency under continuous updates, but the increase includes
delays from the measurement process, preventing separate
quantification of the increase in DUT processing latency.
Update delivery is also affected. Figure 20 shows how the interval
between Serial Notify transmissions changes as the workload
progresses.
+==============+=================================================================+
| Operations | Mean Serial Notify transmission interval (s) |
| per batch | |
+--------------+---------------------+---------------------+---------------------+
| | Batches 1–100 | Batches 100–200 | Batches 200–300 |
+==============+=====================+=====================+=====================+
| 100 | 1.000 | 1.000 | 1.000 |
+--------------+---------------------+---------------------+---------------------+
| 500 | 1.000 | 1.001 | 1.001 |
+--------------+---------------------+---------------------+---------------------+
| 1,000 | 1.008 | 1.277 | 2.087 |
+--------------+---------------------+---------------------+---------------------+
Liu, et al. Expires 6 May 2027 [Page 42]
Internet-Draft ROV Benchmarking November 2026
Figure 20: Serial Notify transmission intervals during continuous
updates.
The Serial Notify transmission intervals in Figure 20 remain close to
the configured one-second interval at 100 and 500 operations per
batch. At 1,000 operations per batch, the interval grows as the test
progresses, exceeding two seconds on average in the final portion.
The RTR source therefore falls behind the submission schedule at the
highest load, alongside the slowdown in state observation.
Appendix B.10. Resource Utilization
This experiment implements the resource-utilization test in
Section 6.6 using the multi-batch VRP update dataset. It replays the
first 60 batches of the workload targeting 100 VRP operations per
batch, with batches submitted at one-second intervals. System and
BIRD CPU and memory are measured during initial synchronization, idle
operation, and updates with convergence. The results are shown in
Figure 21 and Figure 22.
+==================================+=====================+===================+
| Phase | Mean system CPU (%) | Mean BIRD CPU (%) |
+==================================+=====================+===================+
| Initial full-table | 205.497 | 66.479 |
| synchronization | | |
+----------------------------------+---------------------+-------------------+
| Idle steady state | 31.161 | 1.541 |
+----------------------------------+---------------------+-------------------+
| Updates and convergence | 228.661 | 73.150 |
+----------------------------------+---------------------+-------------------+
Figure 21: System and BIRD CPU utilization by experimental phase.
+=================+=============+==============+==============+==============+
| Phase | Mean system | Mean system | Mean BIRD | Mean BIRD |
| | used memory | sampled peak | RSS (MiB) | RSS sampled |
| | (MiB) | (MiB) | | peak (MiB) |
+=================+=============+==============+==============+==============+
| Initial | 4057.007 | 4107.663 | 29.991 | 57.610 |
| full-table | | | | |
| synchronization | | | | |
+-----------------+-------------+--------------+--------------+--------------+
| Idle steady | 6837.546 | 6869.547 | 501.815 | 501.926 |
| state | | | | |
+-----------------+-------------+--------------+--------------+--------------+
| Updates and | 7531.922 | 7633.124 | 503.106 | 503.501 |
| convergence | | | | |
+-----------------+-------------+--------------+--------------+--------------+
Liu, et al. Expires 6 May 2027 [Page 43]
Internet-Draft ROV Benchmarking November 2026
Figure 22: System memory consumption and BIRD RSS by experimental
phase.
Both CPU columns use one fully occupied logical CPU as 100%; system
CPU can exceed 100%. Figure 21 and Figure 22 report means over runs,
and the memory peak column is the mean of each run's sampled phase
peak. Hardware performance counters are not included in this WSL2
campaign. In this test setup, ROV and BGP processing share a single
BIRD process, so their resource usage cannot be measured separately
and is reported for BIRD as a whole.
In Figure 21, BIRD CPU utilization is low during the idle phase and
rises during synchronization and continuous updates. Mean
utilization reaches 73.2% during updates and convergence, compared
with 1.5% while idle. BIRD RSS in Figure 22 remains close to 502–503
MiB between the idle and update phases, while system memory
consumption increases.
Appendix B.11. RTR Session Behavior Tests
These experiments implement Scenarios 1, 3, and 4 of Section 6.7.
Scenarios 1 and 4 use 500k initial VRPs; Scenario 3 uses 50k, 100k,
500k, and 1M. Each input contains 1,000 VRP operations.
Measurements cover RTR recovery, full resynchronization, malformed-
update handling, and the resulting ROV, RIB, and FIB changes.
Scenario 2 is not performed, as explained below. The scenario
triggers and timer origins are summarized in Figure 23.
+======================+==========================+========================+
| Scenario | Trigger | Timer origin |
+======================+==========================+========================+
| 1. Session reset and | Close active RTR | BRIO disconnect |
| recovery | connection | completion |
+----------------------+--------------------------+------------------------+
| 3. Full | Apply updates and send | Cache Reset send time |
| resynchronization | Cache Reset | |
+----------------------+--------------------------+------------------------+
| 4. Normal/malformed | Replay updates; replace | First Prefix PDU send |
| incremental updates | every tenth Prefix PDU | time (relay to DUT) |
| | with type 255 for errors | |
+----------------------+--------------------------+------------------------+
Figure 23: RTR session test scenarios, triggers, and timer origins.
*Scenario 1: Session reset test.*
Liu, et al. Expires 6 May 2027 [Page 44]
Internet-Draft ROV Benchmarking November 2026
+=========+============+===========+==============+============+===================+
| Initial | New | RTR | VRP | ROV | Post-installation |
| VRPs | connection | recovery | installation | completion | RIB/FIB |
| | (ms) | (ms) | (ms) | (ms) | convergence (ms) |
+=========+============+===========+==============+============+===================+
| 500k | 1000.291 | 15568.829 | 15588.969 | 17378.615 | 5692.486 |
+---------+------------+-----------+--------------+------------+-------------------+
Figure 24: RTR session reset and recovery results.
The first four latency columns in Figure 24 start at disconnect
completion. New connection ends when BRIO registers BIRD's new TCP
connection; RTR recovery ends when BIRD's RTR session is observed to
return to the Established state. The last column starts at VRP
installation completion. A new connection is established after
approximately 1.00 s, while RTR recovery and VRP installation
complete after approximately 15.57 s and 15.59 s. ROV completion
follows at 17.38 s, and the post-installation RIB/FIB convergence
interval averages 5.69 s.
*Scenario 2: Cache failover test.*
The cache failover test in Section 6.7 requires a switch from a
primary RTR cache to a secondary cache. In this BIRD setup, both RTR
sessions are active and contribute authorizations to the shared IPv4
and IPv6 ROA tables (r4 and r6). The secondary cache's
authorizations are already available before the primary connection
fails. Disconnecting the primary therefore does not trigger
activation of a standby cache or an identifiable primary-to-secondary
switchover. Measuring continued service from the remaining cache
would not measure the failover required by this scenario, so no
failover latency is reported.
*Scenario 3: Full resynchronization timing.*
Liu, et al. Expires 6 May 2027 [Page 45]
Internet-Draft ROV Benchmarking November 2026
+==========+==========+=============+===================+=============+
| Initial | Full VRP | End-to-end | Post-installation | Final-state |
| VRPs | reload | RIB/FIB | RIB/FIB | consistency |
| | (ms) | convergence | convergence (ms) | |
| | | (ms) | | |
+==========+==========+=============+===================+=============+
| 50k | 160.380 | 4394.699 | 4234.319 | 10/10 |
+----------+----------+-------------+-------------------+-------------+
| 100k | 313.001 | 4986.441 | 4673.440 | 10/10 |
+----------+----------+-------------+-------------------+-------------+
| 500k | 1605.430 | 6794.705 | 5189.275 | 10/10 |
+----------+----------+-------------+-------------------+-------------+
| 1M | 3472.341 | 8575.873 | 5103.532 | 10/10 |
+----------+----------+-------------+-------------------+-------------+
Figure 25: Full VRP resynchronization and BGP convergence.
In Figure 25, mean full VRP reload time increases from 160.380 ms at
50k VRPs to 3472.341 ms at 1M, similar to the full-table
resynchronization results in Figure 4. Mean end-to-end RIB/FIB
convergence increases from 4.395 s to 8.576 s, while mean post-
installation convergence ranges from 4.234 s to 5.189 s. Final-state
consistency checks pass in all ten runs at every scale.
*Scenario 4: Incremental update performance.*
Liu, et al. Expires 6 May 2027 [Page 46]
Internet-Draft ROV Benchmarking November 2026
+=============================+================+=============================+
| Metric | Normal updates | Updates containing |
| | | malformed PDUs |
+=============================+================+=============================+
| VRP installation latency | 126.986 | N/A |
| (ms) | | |
+-----------------------------+----------------+-----------------------------+
| ROV completion latency (ms) | 1586.566 | N/A |
+-----------------------------+----------------+-----------------------------+
| RIB/FIB convergence latency | 6269.012 | N/A |
| (ms) | | |
+-----------------------------+----------------+-----------------------------+
| Error Report received | N/A | 10/10 |
+-----------------------------+----------------+-----------------------------+
| RTR connection closed after | N/A | 10/10 |
| error | | |
+-----------------------------+----------------+-----------------------------+
| Authorization cleanup | N/A | 10/10 |
| completed | | |
+-----------------------------+----------------+-----------------------------+
| Final VRP table correct | 10/10 | 10/10 |
+-----------------------------+----------------+-----------------------------+
| Final ROV states correct | 10/10 | 10/10 |
+-----------------------------+----------------+-----------------------------+
Figure 26: Normal and malformed incremental update results.
In Figure 26, normal updates complete VRP installation, ROV
validation, and RIB/FIB convergence with mean latencies of 126.986
ms, 1.587 s, and 6.269 s, respectively. In all ten malformed-update
runs, the DUT sends an Error Report, closes the RTR connection, and
completes authorization cleanup. Final VRP and ROV checks pass for
both groups.
Authors' Addresses
Libin Liu
Zhongguancun Laboratory
Beijing
China
Email: liulb@zgclab.edu.cn
Nan Geng
Huawei Technologies
Beijing
China
Email: gengnan@huawei.com
Liu, et al. Expires 6 May 2027 [Page 47]
Internet-Draft ROV Benchmarking November 2026
Yanqi Chen
Zhongguancun Laboratory
Beijing
China
Email: chenyq@zgclab.edu.cn
Liu, et al. Expires 6 May 2027 [Page 48]