Auriga pairs Fabric tuning with pre-execution conflict grouping, lifting effective throughput up to six times the default
Synopsis
Auriga splits the effective throughput of Execute-Order-Validate blockchains such as Hyperledger Fabric into raw throughput and transaction success rate, then attacks both with a harness-governed multi-agent configuration search and a scheduler that predicts read-write sets before execution and groups transactions into conflict-free batches. Across three benchmarks and conflict rates up to 70%, effective throughput reaches up to six times the default Fabric configuration and exceeds the strongest standalone baseline; at a 40% conflict rate it holds a 90.2% success rate where the tuning-only Athena stays at 67.8%.
Interpretation
Auriga raises effective throughput, the number of successfully committed transactions per second, from both the configuration side and the conflict side at once, rather than lifting raw throughput alone or repairing conflicts alone. Earlier configuration tuners raised raw throughput while assuming conflict-free workloads, conflict-control systems raised the success rate but repaired conflicts only after execution, and the few joint approaches tuned a small parameter set and relied on kernel changes. At a 40% conflict rate, the tuning-only Athena lifts effective throughput from 205 to 612 but leaves the success rate at 67.8%; Auriga holds a 90.2% success rate at the same conflict rate and reaches an effective throughput of 1,226.
Configuration search is split across three LLM agents that propose configuration hypotheses, drive each candidate to its saturation point, and distill every trial into a reusable experience record. A single free-running agent risks unsafe actions and premature termination, while a hard-coded search procedure leaves the LLM answering stage-specific prompts. A shared harness runtime mediates tool and memory access and holds sole termination authority, granting termination only after verifying exploration depth, group diversification and trend stabilization.
The scheduler predicts each transaction's read-write set before execution and partitions transactions into mutually conflict-free groups bounded by the tuned block capacity. Earlier conflict control parsed read-write sets after execution, so it repaired conflicts reactively instead of preventing them before submission. Offline, an LLM extracts symbolic AST templates from smart contracts and specifies the world-state statistics needed; online, the scheduler instantiates those templates with cached statistics and takes a near-unity-confidence union of execution paths.
Perspective
Evaluation covers five Fabric topologies from 8 to 40 peers, the Smallbank, Token-ERC-20 and IOHeavy contract benchmarks, and conflict rates swept from 10% to 70%, with every tuning method given the same 50-round budget; the scheduling subsystem is compared on an 8-peer topology with MaxMessageCount fixed at 200 and a send rate of 500 tx/s.
The scheduler uses a near-unity confidence threshold, systematically erring toward over-prediction at the cost of some parallelism; prediction coverage is 99.86% overall, and 1.3% of MVCC aborts are still attributed to read-write-set misprediction. Widening the scheduling window from 3,000 to 7,000 transactions raises the success rate from 73% to 99% but denser conflict graphs lower throughput, so 5,000 is the default compromise. The authors list multi-chain and cross-chain environments and adaptive scheduling as future work.
