Introduction
Our earlier Single Asset Vault & Lending Protocol Performance Test Report assessed the original Lending Protocol and Single Asset Vault amendments released in xrpld 3.1.0. This report covers the next version of the protocol.
LendingProtocolV1_1 (XLS-66, XLS-65) is included in xrpld 3.4.0. It introduces three changes:
- cash-basis accounting for loan brokers
- a closed-ended vault type
- stricter vault and loan invariants
All three affect the core lending transactions, LoanSet and LoanPay. We therefore assessed the amendment's performance impact as part of the 3.4.0 release.
Summary
LendingProtocolV1_1 introduces no measurable performance regression. Across about 20 A/B comparisons with the amendment enabled, we observed no regression in:
- throughput
- ledger close time
- per-transaction processing cost
- system resource utilization
| 3.3.0 → 3.4.0 | Lending (3 h) | Mixed payments + lending (3 h) | Endurance (12 h)* |
|---|---|---|---|
| Transactions per cluster | 1.56 M | 1.56 M | 4.8 M |
| TPS | 149.31 → 149.28 | 149.30 → 149.32 | 111.99 → 111.99 |
| Ledger close mean (s) | 3.946 → 3.945 | 4.126 → 4.120 | 3.883 → 3.879 |
| Ledger close P95 (s) | 4.755 → 4.765 | 5.050 → 5.070 | 4.475 → 4.461 |
| Result | No regression | No regression | No regression |
*The endurance test runs a lighter version of the mixed payments + lending workload (about 15% LoanPay) at a fixed 112 TPS to evaluate long-term stability.
Methodology
Test environment. A private XRPL network sized to MainNet specifications:
- 5 validator nodes
- 4 client nodes that accept transactions from the load generators
- 4 load generators
- Every test starts from the same ledger state: about 250,000 funded accounts with pre-built loan brokers, vaults and loans.
A/B comparison. Two identical clusters run side by side under the same load. The baseline cluster runs xrpld 3.3.0. The target cluster runs the new code (individual feature builds, then xrpld 3.4.0) with LendingProtocolV1_1 enabled, so the new vault and invariant checks apply to every lending transaction.
Evaluation. For each test, we compared seven per-ledger metrics between the two clusters, including ledger close time, per-transaction processing time and consensus timing, using statistical tests to separate real changes from normal run-to-run variation.
Workloads.
| Workload | Transaction mix |
|---|---|
| Lending |
LoanSet (40%) and LoanPay (60%) across IOU, MPT and XRP loans |
| Payments | XRP, IOU and cross-currency path payments |
| Mixed payments + lending | Direct and multi-path payments combined with LoanPay, representing a production-like mix |
Results
Feature-Level Testing
We assessed each lending feature individually against 3.3.0, and then all features combined.
| Feature | Workload | TPS | Success rate | Result |
|---|---|---|---|---|
| Cash-basis accounting | Lending | 168 | 99.999% | No regression |
| Closed-ended vault | Lending | 168 | 99.998% | No regression |
| Closed-ended vault | Payments | 149 | 99.996% | No regression |
| All lending features combined | Lending | 168 | 99.999% | No regression |
The payments test confirms that the updated invariant checks, which apply to all transaction types, do not affect non-lending transactions.
Lending Transactions on 3.4.0
| Transaction type | Transactions | Success rate | Avg latency (ms), 3.3.0 → 3.4.0 |
|---|---|---|---|
LoanPay (IOU, MPT, XRP) |
936,708 | 99.999% | 12.51 → 12.52 |
LoanSet (IOU, MPT, XRP) |
622,943 | 99.999% | 25.76 → 23.01 |
12-Hour Endurance
Over 12 hours of sustained load, ledger close time stayed consistent with the 3.3.0 baseline (see summary table). Memory usage followed the same pattern on both versions: it rose during the first few hours as caches warmed up, then held steady. Throughout the test, the two versions stayed within 0.1 GB of each other.
Resource Utilization
| Resource | Result |
|---|---|
| CPU (validators and client nodes) | No regression |
| Memory | No regression |
| Network | No regression |
| Threads, open files, disk I/O | No change |
CPU and memory differences between 3.3.0 and 3.4.0 averaged about 1%, comparable to the run-to-run variation we measured between identical clusters. All nodes operated well below capacity throughout testing.
CPU
Memory
Conclusion
The assessment covered feature-level testing, 3-hour comparisons on two workloads, and 12-hour endurance testing. LendingProtocolV1_1 showed no performance regression in any of these areas.
From a performance standpoint, LendingProtocolV1_1 is ready for MainNet activation.
Top comments (0)