Work

Scalability and Efficiency enhanced based Machine Learning for Hyperledger Fabric and Private Ethereum in Smart Contract Execution

Year: 2026
Type: preprint
Abstract: Blockchain platform selection for enterprise applications increasingly rests on empirical throughput and latency evidence rather than architectural claims alone. This paper proposed a benchmarking work flow based machine learning to compare Hyperledger Fabric v2.5 and private Ethereum (Geth) across four smart contract functions (QueryUser, Crea-5 teUser, IssueMoney, and TransferMoney) under varying network sizes (2–16 peers) and 6 transaction batch sizes (10–1,000), plus an additional 10,000-transaction scalability stress 7 test on Ethereum. Both platforms were independently rebuilt on a single-host Docker 8 deployment, implemented equivalent banking-application smart contracts in Solidity and 9 Go, and reran the full experimental matrix using an in-house benchmarking harness. Our 10 results corroborate the original paper’s central claims: Fabric outperforms Ethereum on 11 write-heavy invoke operations by a factor of roughly 1.4–8.2× depending on network size, 12 while Ethereum’s account-based ledger model yields consistently higher query throughput 13 (1,224 vs. 594 TPS, averaged across all tested batch sizes). This Paper additionally identify 14 a failure mode not emphasized in the original study: Fabric’s multi-key read-modify-write 15 function (TransferMoney) failed under optimistic-concurrency conflicts at every tested 16 batch size and, notably, at a rate that was statistically invariant to peer count. Beyond 17 reproduction, This Paper contributes a supervised machine-learning layer: Random Forest 18 and Gradient Boosting regressors trained on the reproduced dataset to predict through-19 put and latency directly from network configuration. Under leave-one-peer-count-out 20 cross-validation, the surrogate model predicts per-transaction latency with high fidelity 21 (R 2 = 0.887) but throughput far less reliably (R 2 = 0.478), a gap attribute to the discontinu-22 ous, contention-driven nature of the throughput data itself. This paper discuss deviations 23 from the original study’s setup and their implications for the generalizability of both papers’ 24 findings.
Institution: Minia University
Language: English
Open Access status: green
OpenAlex ID: W7225994497
Cites: 5
Cited by:
Related to: 10