Optimal Sharding for Scalable Blockchains with Deconstructed SMR

Abstract

Sharding is proposed to enhance blockchain scalability. However, a size-security dilemma where every shard must be large enough to ensure its security constrains the efficacy of individual shards and the degree of sharding itself. Most existing sharding solutions therefore rely on either weakening the adversary or making stronger assumptions on network links. This paper presents Arete, an optimally scalable blockchain sharding protocol designed to resolve the dilemma based on an observation that if individual shards can tolerate a higher fraction of (Byzantine) faults, we can securely create smaller shards in a larger quantity. The key idea of Arete, therefore, is to improve the security resilience/threshold of shards by dividing the blockchain's State Machine Replication (SMR) process itself. Similar to modern blockchains, Arete first decouples SMR in three steps: transaction dissemination, ordering, and execution. However, unlike other blockchains, for Arete, a single ordering shard performs the ordering task while multiple processing shards perform the dissemination and execution of blocks. As processing shards do not run consensus, each of those can tolerate up to half compromised nodes. Moreover, the SMR process in the ordering shard is lightweight as it only operates on the block digests. Second, Arete considers safety and liveness against Byzantine failures separately to improve the safety threshold further while tolerating temporary liveness violations in a controlled manner. Apart from the creation of more optimal-size shards, such a deconstructed SMR scheme also empowers us to devise a novel certify-order-execute architecture to fully parallelize transaction handling, thereby improving the performance of sharding systems. We implement Arete and evaluate it on a AWS environment by running up to 500 nodes, showing that Arete outperforms the state-of-the-art sharding protocols.