Post-Quantum Signature Experiments and Migration Considerations for the Resource Public Key Infrastructure (RPKI)

7.1. Parallel Publication

RFC 6916 specifies a top-down transition in which a parent CA adopts support for the Next Suite before its children. During phases 2 and 3, corresponding Current Suite and Next Suite product sets are maintained in parallel. RFC 6916 deliberately avoids mixed-suite CA certificates: a CA certificate signed using one suite does not carry a subject key associated with another suite.¶

The Parallel Publication experiment publishes corresponding Current Suite and Next Suite products separately. It does not define new payload encodings for manifests, ROAs, or CRLs. As described in Section 6 of [RFC6916], RPs can retrieve product sets under both algorithm suites and compare their outputs for testing. This experiment performs that comparison using the resulting VRP sets and CCR [I-D.ietf-sidrops-rpki-ccr].¶

The experimental harness needs a mapping between corresponding products for measurement and debugging. Such a mapping can be derived from the publication point structure, object names, CA hierarchy, or an implementation-specific record. It is not proposed as a new on-wire RPKI object.¶

7.2. Mixed Tree

A parent using the Current Suite signs a child CA certificate whose subject public key belongs to the Next Suite. The child then issues its certificates, CRLs, and CMS signed objects using the Next Suite without first requiring the parent to migrate its own CA key and product set. Production issuance would still require the parent to process the child's Next Suite certificate request and proof of possession. This experiment constructs the child CA certificate directly; it has not implemented that provisioning exchange.¶

The Mixed Tree migration approach is independent of the selected signature suite. The evaluated repository instantiates the Next Suite with the Composite ML-DSA configuration evaluated in this revision. This construction is not permitted by the RFC 6916 transition procedure and is evaluated only as an experimental alternative under test TALs.¶

Unlike the RFC 6916 procedure, a switched subtree does not maintain corresponding Current Suite and Next Suite products. This avoids duplicate repository content and the publication, configuration, and rollover work needed to keep two product sets aligned. It also permits subtrees to move independently after the necessary parent and RP support is available, rather than waiting for every ancestor to migrate its own CA key and product set.¶

The trade-off is compatibility. Once a subtree switches, an RP that does not support the Next Suite cannot validate it through the certification path. Mixed Tree deployment therefore replaces the parallel legacy hierarchy with a requirement for sufficient RP support before each subtree is migrated. The complete certification path remains dependent on Current Suite certificate signatures above the migration boundary.¶

The experiment uses test repositories and test TALs. It does not define or authorize a production transition procedure.¶

8.1. Current Suite

The Current Suite baseline used by this experiment is RSA PKCS #1 v1.5 with SHA-256 as specified by RFC 7935. This document does not set a production transition timetable.¶

8.2. Composite Configuration Evaluated in This Revision

The composite signature construction used by this experiment is defined in revision 19 of [I-D.ietf-lamps-pq-composite-sigs]. The evaluated configuration is id-MLDSA65-ECDSA-P256-SHA512 (OID 1.3.6.1.5.5.7.6.45), whose identifier and component-algorithm combination are specified by that document. It combines ML-DSA-65 with ECDSA P-256 and requires both component signatures to validate. The corresponding CMS use is specified by [I-D.ietf-lamps-cms-composite-sigs]. The choice of this composite configuration is an experimental input, not a recommendation for production deployment.¶

The object experiment applies [I-D.ietf-lamps-pq-composite-sigs] to composite public keys and certificate signatures, and [I-D.ietf-lamps-cms-composite-sigs] to composite CMS SignedData. The public reference implementation has generated complete composite X.509 certificates, CRLs, ROAs, and manifests and has validated a Mixed Tree with an RSA parent and a Composite child using experimental rpki-client and Routinator extensions. The two RP processing paths share one OpenSSL and Composite provider backend, so this is not independent cryptographic interoperability evidence.¶

Experiments with the composite suite are confined to isolated repositories under test TALs. Experimental objects, production repositories, production keys, and production TALs are deliberately kept separate.¶

8.3. Additional Candidate Suites

ML-DSA-44 is an additional experimental configuration. It produces smaller public keys and signatures and may sign and verify faster, which matters in a system where RPs repeatedly synchronize repository state and validate large numbers of objects. This revision uses ML-DSA-65 in the composite configuration as an explicit experimental choice, not because the RPKI requires NIST security Category 3. The Rationale for the Experimental Configurations section discusses this choice and its counterarguments.¶

ML-DSA-87 is included as a higher-security comparison candidate. It is more conservative than ML-DSA-65 but carries correspondingly larger size and performance costs.¶

SLH-DSA-SHAKE-128s and SLH-DSA-SHAKE-192s are included for cryptographic-diversity comparison. Their signature algorithm, PKIX, and CMS specifications are defined by [FIPS205], [RFC9909], and [RFC9814]. They are not proposed as the initial suite in this revision because their signature sizes are substantially larger than those of ML-DSA in the evaluated configurations.¶

FN-DSA (Falcon), MAYO, and SNOVA are additional candidates for future evaluation. They are outside the present evaluation until stable PKIX and CMS profiles are available and referenced by a future revision or separate document. FN-DSA in particular is discussed further in the Rationale for the Experimental Configurations section, because its compact signatures make it an attractive candidate for the RPKI's bulk validation model.¶

Other Composite ML-DSA combinations specified by LAMPS remain candidates for comparison. Changing the component pair would change object sizes and cryptographic costs, but would not determine whether a deployment uses Parallel Publication or Mixed Tree migration. RP support for a selected suite remains a deployment prerequisite.¶

8.4. Classical Reference Points

To relativize the cost of PQC candidates, this document uses two compact classical algorithms as non-normative reference points: ECDSA P-256 with SHA-256 [FIPS186-5], which is already used for BGPsec UPDATE signatures [RFC8608], and Ed25519 [RFC8032]. Neither is a CRQC-resistant algorithm, and neither is proposed here as an RPKI suite. They provide compact classical reference points for signature and key sizes in the classical setting: the deployed RSA-2048 profile is itself several times larger than these curves, and PQC candidates should be compared against both baselines rather than against RSA alone.¶

8.5. Size Model Inputs

The table below records the values supplied to the first-order size model. The representations are deliberately mixed: some rows use raw public keys or fixed-width signatures, while others use representative DER values or a maximum encoded signature size. They are therefore model inputs, not a uniform comparison of SubjectPublicKeyInfo or signature encodings.¶

"Cat." is the NIST security category; "n/a" marks classical reference algorithms. The RSA-2048 key value is a representative DER RSA public-key value used by the model. The P-256 value is an uncompressed point, and the Ed25519, ML-DSA, FN-DSA, and SLH-DSA values are raw public keys. The ECDSA value is a representative DER signature size, while FN-DSA records a maximum. These choices are documented so that later models can replace them with consistently measured RPKI object encodings.¶

Beyond static sizes, the evaluation considers the following dimensions: certificate and CRL size under the RPKI profile; CMS signed object size; signing and verification cost under the RP workload; impact on repository size and on distribution via RRDP, rsync, and Erik; CA key rollover and publication cycle impact; HSM support; and standardization and implementation maturity. The dimensions backed by measurements in this revision are identified in Appendix A.¶

Two qualitative observations from the preliminary evidence inform the rationale in the next section. First, RP workload is verification dominated: an RP verifies repository objects but signs nothing as part of validation, and the deployed RSA profile is exceptionally cheap at verification, so increases in verification cost are particularly relevant to RP operation. Second, first-order size models place ML-DSA-65 near a fourfold repository size increase over the RSA baseline, ML-DSA-44 near threefold, and FN-DSA-512 well below twofold.¶

Measurements supporting these observations, together with their conditions, caveats, and the list of dimensions not yet backed by confirmed measurements, are collected in Appendix A and are maintained in reproducible form by the experimental harness [pqc-rpki-lab]. Measured values are implementation and environment dependent and are not protocol requirements.¶

8.6. Rationale for the Experimental Configurations

This experiment evaluates a composite suite because a global RPKI migration may benefit from not depending exclusively on a newly deployed PQC algorithm. The LAMPS construction accepts a composite signature only when both component signatures validate. Under the assumptions stated by that construction, its EUF-CMA guarantee is retained when at least one component remains EUF-CMA secure and the prehash remains collision resistant.¶

This guarantee has a limit. A CRQC defeats the ECDSA component, so security against a quantum adversary still depends on ML-DSA-65 remaining secure. The composite suite also does not protect against failures shared by both components or by the combiner, encoding, key management, or validation implementation. It does not in general preserve strong unforgeability when only one component has that property. An implementation defect in one component can be tolerated only when the other component, the combiner, encodings, key handling, and the verification of all component signatures remain unaffected.¶

ML-DSA-65 is used as the PQC component of the composite configuration evaluated in this revision because it has a final FIPS signature specification [FIPS204], corresponding PKIX [RFC9881] and CMS [RFC9882] algorithm identifier specifications, and implementations available in the software environment used by this experiment. This revision also chooses Category 3 as an experimental point for examining the cost of a larger security margin in a system in which another global algorithm migration could take multiple years. This is not an RPKI requirement. It is not used because it is the smallest or fastest possible signature algorithm; it is neither.¶

The choice between ML-DSA-44 and ML-DSA-65 is genuinely open. The argument for ML-DSA-65 is conservatism: the RPKI is a single global system, algorithm migrations in it are slow and expensive, and a larger security margin reduces the probability of needing another migration. The argument for ML-DSA-44 is that a structural cryptanalytic break of module lattices would likely affect all ML-DSA parameter sets, so the extra category mainly protects against gradual erosion of concrete security estimates rather than against a qualitative break; under that view, the roughly 25-35% smaller keys and signatures of ML-DSA-44, or a composite configuration using ML-DSA-44, may be a better use of the size budget [Doesburg2025]. This document uses ML-DSA-65 in the composite configuration measured in this revision, keeps ML-DSA-44 in the comparison, and records the parameter-set choice as a question for further work.¶

ML-DSA-87 provides a higher-security comparison point, but its size and performance costs make it less attractive in the evaluated repository model.¶

FN-DSA (Falcon) is an attractive candidate on size and performance grounds: its signatures are roughly one fifth the size of ML-DSA-65 signatures, and both published RPKI analysis [Doesburg2025] and repository-scale redesign work [pqRPKI] identify Falcon as the compact lattice option. This document nevertheless treats FN-DSA as an additional configuration for future evaluation rather than one measured in the composite experiment, for the following reasons:¶

• This document does not profile FN-DSA because it does not reference a final FN-DSA standard together with stable PKIX and CMS profiles.¶

• Side-channel-resistant FN-DSA implementations and HSM support require separate implementation and deployment evaluation.¶

• Availability of FN-DSA in the platforms that RPKI CAs, RIRs, HSM vendors, and validator implementations actually use is a separate question from the algorithm's intrinsic merits. This experiment does not evaluate production RPKI CA or RP support for FN-DSA.¶

These are reasons to sequence the evaluation, not to dismiss the algorithm. FN-DSA remains in the comparison set as the compact signature candidate, and the conditions under which it should be examined further are recorded as questions for further work.¶

Algorithm selection for the RPKI cannot be based on software benchmarks alone. HSM support for a candidate algorithm is also a deployment prerequisite for CAs that protect their signing keys in HSMs.¶

9.1. Certificate and CRL Encoding

The composite certificate and CRL experiment follows the encodings in [I-D.ietf-lamps-pq-composite-sigs]. Under that assumption, a composite subject carries id-MLDSA65-ECDSA-P256-SHA512 in the SPKI AlgorithmIdentifier with absent parameters. A certificate or CRL signed by the composite issuer uses that identifier in its signatureAlgorithm field. A child CA certificate signed by a Current Suite issuer instead retains the issuer's Current Suite signatureAlgorithm while carrying the composite identifier in the subject SPKI.¶

The same composite SPKI and signature encodings are assumed for resource certificate requests and their proof-of-possession signatures. Pure ML-DSA certificates and CRLs used for component measurements follow [RFC9881]; they provide component-level evidence and are not proposed as a production Next Suite. These request and proof-of-possession encodings are not exercised by the implementation in this revision.¶

The generated RPKI test objects preserve the existing keyUsage constraints. CA certificates carry keyCertSign and cRLSign, while EE certificates for signed objects carry digitalSignature and are not used as CA certificates. The experiment does not change resource extension semantics, the certificate policy OID, certificate path validation, manifest processing, or CRL processing.¶

9.2. CMS Signed Object Encoding

The composite CMS experiment combines [I-D.ietf-lamps-cms-composite-sigs] with the RPKI signed object template in [RFC6488], as updated by [RFC9589]. The assumed SignerInfo signatureAlgorithm is id-MLDSA65-ECDSA-P256-SHA512 with absent parameters. The SignedData digestAlgorithms set includes id-sha512, and the SignerInfo digestAlgorithm is id-sha512. The parameters are absent in both AlgorithmIdentifiers, and the message-digest signed attribute contains the SHA-512 digest of the eContent.¶

The experiment retains the signedAttrs restrictions from the existing RPKI signed object profile: one content-type attribute, one message-digest attribute, and one signing-time attribute, with no binary-signing-time or other signed attributes. The eContentType, eContent, and object-specific validation rules remain unchanged.¶

The implementation generated and validated complete Composite ROA and manifest SignedData using these encodings, together with the certificates and CRLs required for repository validation. Experimental rpki-client and Routinator extensions processed the resulting repository. Interoperability with a cryptographic implementation independent of the shared OpenSSL and Composite provider remains open work.¶

9.3. Manifest Processing During Migration

A manifest covers the products of one CA instance at one publication point, as specified by [RFC9286] and updated by [RFC9981]. The manifest is signed with a one-time-use EE certificate issued by that CA. Its fileList contains the certificates issued and published by that CA, the CA's current CRL, and signed objects whose embedded EE certificates were issued by that CA.¶

The relevant RP check is therefore issuer and publication-scope consistency, not equality between the manifest signing key and product keys. An RP validates the manifest EE certificate under the associated CA, verifies each listed certificate, CRL, or signed object under that same CA instance as required by its object profile, and checks the publication point, file name, and file hash according to [RFC9286]. A shared publication point can contain products from multiple CA instances during key rollover, but each manifest covers only its associated CA instance.¶

Mixed Tree migration and composite signatures do not change these checks. This document therefore introduces no additional requirement for the manifest EE key to equal a key used by a listed product, and it does not weaken the existing RP checks that bind every listed product to the manifest's CA scope.¶

11.1. Impact of Larger Objects

Changing the signature suite affects more than cryptographic processing. Larger certificates, CRLs, manifests, and signed objects increase the bytes transferred during initial repository synchronization and in updates that contain those objects. The captured 1,000-ROA Composite repository occupied 9,797,552 object bytes, compared with 1,768,736 bytes for its RSA rollback state, a 5.54-fold increase. Transport scalability is therefore a deployment constraint that needs evaluation independently of algorithm correctness and RP validation time.¶

11.2. RRDP, rsync, and Erik Synchronization

RRDP [RFC8182] distributes snapshots and a journal of deltas. An RP obtains a complete snapshot when the notification file no longer offers a contiguous delta chain from the RP's serial, or when the session identifier changes. A polling RP can receive intermediate publication states that it does not ultimately use. An rsync synchronization first exchanges repository metadata and then transfers changed files; even an unchanged repository therefore has a cost that grows with the file list.¶

Erik Synchronization [I-D.ietf-sidrops-rpki-erik-protocol] is an RPKI- specific replication protocol using Merkle trees, content-addressable naming, and HTTP. An RP compares ErikIndex and ErikPartition objects and fetches selected objects by hash. Snapshot and segment prefetching can reduce the request cost of cache bootstrap and catch-up. This document does not specify Erik or require it for PQC deployment; it evaluates Erik alongside RRDP and rsync as a candidate response to repository expansion.¶

11.3. Transport Experiment Boundary

Appendix A compares initial synchronization, an unchanged repository, one ROA replacement, and 10% ROA churn for RSA-2048, pure ML-DSA-65, and the evaluated Composite suite. It combines an actual local rsync run with calculated RRDP and Erik response-body sizes over the same deterministic corpus, calibrated to the measured object counts and repository sizes. The corpus preserves measured object counts and repository sizes but is not a cryptographically valid RPKI repository. The comparison is consequently evidence about byte growth, transferred bytes, and request counts, not a production-network throughput result.¶

12.1. Results Summary

OpenSSL 3.6.2 and the evaluated Composite provider generated complete certificate, CRL, manifest, and ROA sets for pure ML-DSA-65, Composite ML-DSA, and a Mixed Tree with an RSA parent and a Composite child. Experimental rpki-client and Routinator extensions validated each repository and produced the same two VRPs as the RSA baseline. An experimental Krill extension then created a Composite child below an RSA parent, published and replaced ROAs, and rolled the child back to RSA. Both experimental RPs derived the expected VRPs at each stage.¶

The primary repeated cryptographic-operation benchmarks measure key generation, signing, and verification without X.509, CMS, repository transfer, or RP processing. Those rows use ten independent runs and give the median and sample standard deviation. Additional Composite ML-DSA operation measurements and their conditions are identified separately in Appendix A. Repository ratios are separate synthetic estimates derived from explicit object counts and measured key and signature sizes. Appendix A records the complete methods, results, and limitations.¶

The two RP extensions use the same OpenSSL and Composite provider backend. The experiment therefore demonstrates two RP processing paths, but not independent cryptographic interoperability or production readiness.¶

12.2. Relying Party Evaluation

The experiment extended rpki-client and Routinator to recognize the id-ml-dsa-65 and id-MLDSA65-ECDSA-P256-SHA512 identifiers and to delegate their cryptographic operations to OpenSSL providers. Each RP processed four complete repositories: the RSA baseline, pure ML-DSA-65, Composite ML-DSA, and a Mixed Tree with an RSA parent and a Composite child. For every repository, the test covered certificate-path and CRL validation, manifest and CMS validation, ROA processing, and VRP production.¶

The Mixed Tree test additionally verified that the issuer signature algorithm and subject SPKI algorithm are processed independently. The RSA parent's public key verifies the signature on the child CA certificate, while the Next Suite public key carried in that certificate's SPKI is used to validate signatures issued by the child.¶

When an experiment validates both Current Suite and Next Suite products, it separately compares their validated outputs. For ROAs, the comparison uses equality of the canonical VRP sets by prefix, maxLength, and origin AS. When both runs use the same CCR version and hash algorithm, equal ROAPayloadState hashes indicate equality of those canonical VRP sets under the collision-resistance assumption of the CCR digest [I-D.ietf-sidrops-rpki-ccr]. The experiment parsed actual rpki-client CCR DER for the RSA, pure ML-DSA-65, Composite, and Mixed Tree repositories and recomputed every embedded collection hash. All four ROAPayloadState hashes were equal, while ManifestState differed as expected. TrustAnchorState is compared separately and also differed. This result uses one RP implementation that produced CCR output. ROAPayloadState does not preserve per-VRP certificate-chain or publication provenance, so comparing the certificate chain and publication source associated with each VRP requires additional experiment records. Differences in validated outputs are recorded as an experimental result rather than silently merged.¶

This document does not require routers to support PQC. Routers receive validated payloads through RTR or local export formats, and the semantic content of that output is intended to be unchanged by the algorithm migration.¶

16.1. Algorithm Configurations

• Whether ML-DSA-44, ML-DSA-65, or another candidate provides an appropriate balance of object size, signing and validation cost, security margin, and implementation availability for a future standards-track profile.¶

• Under what conditions FN-DSA should be promoted from a future evaluation candidate: completion of the FN-DSA standard, stable PKIX/CMS conventions, evidence of side-channel-resistant implementations, and HSM availability.¶

16.2. Migration Design

• Under what operational and compatibility conditions Parallel Publication or Mixed Tree migration provides an acceptable migration path.¶

• How RP readiness can be measured reliably before migrating a subtree in a production Mixed Tree deployment.¶

• How to define a transition timetable and readiness metrics, and whether that work should update or replace RFC 6916.¶

• How the EE subject public-key algorithm, the CA signature on the EE certificate, and the CMS signature algorithm should be related under a future Next Suite, including whether a null-signature construction should replace the current one-time-use EE model.¶

• How provisioning and publication software will roll the BPKI trust anchors and EE certificates used for existing relationships, including relationships established through [RFC8183], before those protocols depend on a PQC algorithm. The procedure needs overlap, rollback, and recovery behavior and can be prepared independently of the final RPKI object-signature algorithm choice.¶

16.3. Repository Distribution

• Whether RRDP and rsync remain operationally sufficient at projected PQC repository sizes, topology, polling intervals, and churn rates, or whether deployment requires a more selective mechanism such as Erik Synchronization.¶

16.4. Operational Readiness

• Which PQC signature algorithms RIR CA teams and their HSM vendors plan to support, on what firmware, API, certification, and deployment timelines.¶

A.1. Reproducibility Metadata

The evidence snapshot cited by this revision is the GitHub release draft-yoshikawa-sidrops-pqc-rpki-02 of [pqc-rpki-lab]. The repository contains the scripts, raw outputs, and additional environment metadata used for the measurements in this appendix.¶

The cryptographic-operation measurements were performed on an Apple M4 running macOS 26.5.2 with OpenSSL 3.6.2 and liboqs 0.15.0. The primary algorithm benchmark was compiled with optimization level -O2, while the Composite ML-DSA benchmark was compiled with -O3. The latter implements revision 19 of [I-D.ietf-lamps-pq-composite-sigs].¶

The X.509, CMS, and RP experiments used OpenSSL 3.6.2 and the experimental Composite implementation recorded in the evidence snapshot. End-to-end validation was also exercised on a 12-vCPU x86-64 host using rpki-client 9.8. The RSA, pure ML-DSA-65, Composite, and Mixed Tree repositories all produced the expected VRPs.¶

A.2. Repeated Cryptographic Operation Timing

This experiment compares the raw signing and verification cost of the evaluated algorithms independently of X.509, CMS, repository transfer, and RP processing. In particular, verification performance is relevant to RPs, which repeatedly validate large numbers of repository objects.¶

The table reports the median and sample standard deviation across ten repetitions of 1,000 operations on a fixed 32-byte message. Values are wall-clock seconds per 1,000 operations.¶

The measurements show that RSA-2048 is substantially faster at verification than the other evaluated algorithms in this environment. ML-DSA verification cost increases with the parameter set, with ML-DSA-65 requiring about 7.5 times the RSA verification time and ML-DSA-87 about 11.8 times. Signing shows a different pattern: ML-DSA-44 is faster than RSA-2048, while ML-DSA-65 and ML-DSA-87 are somewhat slower.¶

These measurements isolate cryptographic operations and therefore do not represent complete RP-validation performance. Their main implication for the RPKI is that algorithm selection should consider verification cost separately from signing cost, because RP workloads are verification dominated.¶

The benchmark also covers 512-byte, 2-KiB, and 8-KiB messages and records key-generation timing, variance, and process peak RSS. Those raw results remain in the evidence snapshot rather than being duplicated here.¶

A.3. Composite ML-DSA Operation Measurements

This experiment measures the additional cryptographic cost of combining ML-DSA with a classical signature algorithm in the Composite construction. It compares signing and verification time, public-key size, signature size, and the resulting synthetic repository-size ratio across the evaluated Composite configurations.¶

The following single-run measurements execute 100,000 signing operations and 100,000 verification operations for each configuration. They include message-representative construction, ML-DSA context binding, both component signature operations, raw key and signature concatenation, and verification of all component signatures. The ML-DSA-87 configuration uses P-384 because revision 19 of [I-D.ietf-lamps-pq-composite-sigs] does not define ML-DSA-87 with P-256.¶

The measurements show that both cryptographic cost and object size increase with the ML-DSA parameter set. For the evaluated ML-DSA-65 + P-256 configuration, the classical component adds relatively little to the repository-size ratio compared with ML-DSA-65 alone, while verification must still perform both component checks.¶

These measurements isolate the Composite cryptographic operations. They exclude key generation, file I/O, X.509, CMS, RP processing, and HSM latency, and the repository ratios are synthetic model outputs rather than complete-repository measurements.¶

A.4. Measured Certificate and CRL Sizes

This experiment measures the size impact of the evaluated signature algorithms when applied to RPKI-profiled certificates and CRLs. It provides measured object sizes for comparison with the repository-size models and transport experiments in the following sections.¶

RFC 6487-profiled certificates, including RFC 3779 resource extensions, and CRLs were generated with OpenSSL 3.6.2. The RSA, P-256, Ed25519, ML-DSA, and SLH-DSA rows use the OpenSSL default provider. The FN-DSA-512 row uses the experimental provider described below.¶

The P-256 and Ed25519 rows use the same RFC 6487 structure and resource extensions as the other rows, but are classical comparison algorithms rather than RFC 7935 suites. The FN-DSA-512 row uses the experimental Falcon-512 OID and encoding from oqs-provider 0.11.0-rc1 with liboqs 0.15.0; it is a measured experimental encoding, not a final FN-DSA PKIX profile. Falcon signatures are variable length, so its certificate and CRL sizes can vary between runs.¶

The measurements show that certificate and CRL size varies substantially among the candidate algorithms. ML-DSA increases object size by several times relative to RSA-2048, while the evaluated SLH-DSA configurations produce substantially larger objects. FN-DSA-512 is comparatively compact in this experiment, although its measured encoding is not a final FN-DSA profile. These differences directly contribute to the repository-size and transport costs evaluated in the following sections.¶

A.5. Synthetic Repository Size Model

This model estimates how the measured key and signature sizes translate into repository growth when applied across a representative set of RPKI objects. It is intended as a first-order comparison between candidate algorithms rather than as a prediction of the size of the global RPKI.¶

The synthetic corpus contains 10 CA certificates, 100 EE certificates, 10 CRLs, 10 manifests, and 100 ROAs. Its base payload assumptions are 1500, 1500, 600, 1500, and 200 bytes, respectively.¶

For certificates, the model adds both a public-key input and a signature input to the base payload. For CRLs, manifests, and ROAs, it adds the signature input. The resulting object sizes are multiplied by the corresponding object counts and summed. The repository ratio is then normalized to the RSA-2048 baseline. Transport encoding and cache overhead are not included.¶

As a consistency check, the same formula was applied to the standalone repository object counts: one CA certificate, two EE certificates, one CRL, one manifest, and one ROA. The measured medians are the four required products of the standalone repositories generated by the end-to-end experiment identified in the Reproducibility Metadata section. The error below is (predicted - measured) divided by measured.¶

For these counts, the model predicts a Composite-to-RSA ratio of 3.62, whereas the measured ratio is 5.96. The difference results primarily from the fixed base-payload assumptions, which overestimate the small RSA repository.¶

The model shows that repository growth is strongly dependent on signature and public-key size. ML-DSA-44 and ML-DSA-65 produce approximately three- and four-fold increases over the RSA baseline in this corpus, while the evaluated Composite configuration is close to ML-DSA-65 because the additional P-256 component is small relative to the ML-DSA component. SLH-DSA produces substantially larger estimates, whereas FN-DSA remains comparatively compact.¶

These ratios depend on the assumed object mix and base payload sizes and must not be treated as measured repository-wide growth factors. The measured repository experiments below provide more representative evidence for the configurations that were implemented.¶

A.6. Controlled Repository Scale Measurements

This experiment examines how object generation and RP validation behave as the number of RPKI objects and child CAs increases. A public Routinator cache snapshot was used only as a reference for the scale of the deployed RPKI; the performance measurements themselves use controlled repositories.¶

The reference snapshot, collected with Routinator 0.15.2 using RRDP, contained 550,210 objects across 54,960 publication points and produced 980,019 VRPs. The ARIN trust anchor was unavailable during collection, so these values do not represent the complete global RPKI.¶

A controlled Krill experiment used one RSA parent and one Composite child. Complete repository generation was repeated 30 times for 1, 10, and 100 ROAs and 10 times for 1,000 ROAs. Each repository size was also validated 100 times with a fresh validator cache using the experimental rpki-client and Routinator implementations. Both RPs produced the expected VRPs in every run.¶

The captured 1,000-ROA Composite repository contained 1,008 files occupying 9,797,596 bytes. Its RRDP snapshot was 13,145,809 bytes uncompressed and 9,190,012 bytes with deterministic gzip.¶

The same repository was used to compare validation with a fresh cache, an unchanged repository, and a one-ROA update, with 30 repetitions per RP and condition. The update changed the ROA, manifest, and CRL. rpki-client wall time medians were 0.86, 0.86, and 0.85 seconds, respectively; Routinator medians were 2.26, 1.99, and 2.30 seconds. The OS page cache was not controlled, so these results do not demonstrate a general incremental- validation speedup.¶

A separate synthetic topology used one RSA parent and 100 Composite child CAs, each publishing one ROA. The repository contained 403 objects and produced 100 VRPs with both experimental RPs. Removing one child publication point reduced the result to 99 VRPs without affecting the remaining children.¶

The results show that the prototype continues to produce the expected VRPs as the controlled repository grows, while repository generation cost increases substantially with the number of ROAs. The cache experiment does not show a consistent reduction in validation time after a small update, and the 100-child experiment confirms that failure of one child publication point does not invalidate sibling results. These experiments characterize controlled scaling behavior; they do not measure performance on the complete public RPKI or over production repository transport.¶

A.7. Repository Transport Measurements

This experiment evaluates how larger RPKI objects affect repository synchronization under rsync, RRDP, and Erik Synchronization. It compares initial synchronization, an unchanged repository, a one-ROA update, and 10% ROA churn for RSA-2048, ML-DSA-65, and the evaluated Composite configuration.¶

The experiment reused the measured 1,000-ROA repository size and object count. Each state contained 1,008 files: two certificates, three CRLs, three manifests, and 1,000 ROAs. The RSA and Composite totals were taken from captured repositories, while the ML-DSA-65 total was derived from measured certificate, CRL, manifest, and ROA sizes.¶

For each transition, five local rsync runs started from the same cache state. RRDP values are calculated from notification, snapshot, and delta response bodies. Erik values use a simplified single-partition model with one ErikIndex, one ErikPartition, and the required objects fetched by hash. Protocol headers, TLS, connection setup, and Compression Dictionary Transport [RFC9842] are not included.¶

The simplified Erik model required 1,010 requests for initial synchronization, one for an unchanged repository, five for a one-ROA update, and 104 for 10% churn. With snapshot prefetch, the modeled initial object transfer was consolidated into one bulk response of 1,768,736 bytes for RSA, 9,623,780 bytes for ML-DSA-65, and 9,797,552 bytes for Composite.¶

The results show that larger signatures substantially increase transfer volume during initial synchronization and repository churn, while the cost of an unchanged repository remains small for RRDP and the modeled Erik case. Selective synchronization also avoids retransmitting most unchanged object data in the update scenarios. Repository distribution therefore becomes a more significant consideration as RPKI object sizes increase.¶

These results do not establish the relative production performance of the three protocols. The rsync measurements were local, while the RRDP and Erik results are response-body models rather than measurements from production servers. Network latency, HTTP behavior, compression effects, and retrieval of obsolete intermediate states remain to be evaluated.¶

A.8. Open Measurement Tasks

The following dimensions are not yet backed by confirmed measurements and are deliberately recorded as open tasks rather than numbers:¶

• CA key rollover, publication cycle, and full validation across the evaluated configurations over a multi-CA corpus modeled on the public RPKI that can be re-signed under each evaluated configuration. The controlled single-child and synthetic 100-child results do not reproduce public RPKI topology.¶

• Production RRDP, rsync, and Erik transfer, churn, polling-interval, and cache behavior. The generated bodies and local runs do not measure WAN behavior or obsolete intermediate-state transfer.¶

• Long-running validator memory and cache growth. The current results record process peak RSS for bounded executions only.¶

• HSM performance and support.¶