Operation-level Concurrent Transaction Execution for Blockchains

TL;DR

This paper introduces an operation-level concurrency control algorithm for blockchains that allows conflict-free operations to execute in parallel, significantly improving throughput and speedup over traditional serial execution methods.

Contribution

The paper presents a novel operation-level concurrency control algorithm that enhances blockchain transaction throughput by resolving conflicts at the operation level with a redo mechanism.

Findings

Achieves an average speedup of 4.28× in transaction processing.

Further accelerates execution by 7.11× when combined with state prefetching.

Demonstrates effectiveness on real-world Ethereum blocks.

Abstract

Despite the success in various scenarios, blockchain systems, especially EVM-compatible ones that serially execute transactions, still face the significant challenge of limited throughput. Concurrent transaction execution is a promising technique to accelerate transaction processing and increase the overall throughput. Existing concurrency control algorithms, however, fail to obtain enough speedups in real-world blockchains due to the high-contention workloads. In this paper, we propose a novel operation-level concurrency control algorithm designed for blockchains. The core idea behind our algorithm is that only operations depending on conflicts should be executed serially, while all other conflict-free operations can be executed concurrently. Therefore, in contrast to the traditional approaches, which block or abort the entire transaction when encountering conflicts, our algorithm introduces a redo phase to resolve conflicts at the operation level by re-executing conflicting operations only. We also develop a set of data dependency tracking mechanisms to achieve precise identification and speedy re-execution for conflicting operations. We implement an open-source prototype based on Go Ethereum and evaluate it using real-world Ethereum blocks. The evaluation results show that our algorithm achieves an average speedup of 4.28$\times$. If combined with state prefetching techniques, our approach can further accelerate the transaction execution by 7.11$\times$.