DATS: Scaling Out Blockchain With Deep-Learning-Powered Dynamic Sharding

Abstract

Sharding is a promising solution to deal with the low throughput and poor scalability issues of the blockchain system. It horizontally scales the blockchain by separating the network into several sub-networks known as shards and allowing for parallel transaction processing across shards. However, sharding introduces a significant number of cross-shard transactions, which severely decreases system performance. Previous research has focused on static sharding, which assigns accounts with frequent trading activity to the same shard once and for all. Unfortunately, accounts’ trading behavior is not constant; the initial account sharding becomes ineffective as the blockchain system runs over time. In this paper, we proposed DATS, a novel dynamic sharding scheme designed to facilitate cost-efficient account adjustments in blockchain. We propose a transaction-driven account model where each account consists of multiple sub-accounts in different shards. By resetting the status of sub-accounts via a migration transaction, we can make account adjustments at a low cost and convert cross-shard transactions into intra-shard transactions. Moreover, we develop a time-series prediction model to determine the shards where accounts should be located by forecasting the volume of cross-shard transactions from each sender. We also propose a mechanism for handling hot shards by redirecting transactions to less burdened shards while maintaining workload balancing. We validate our proposed approach through extensive experiments using Ethereum data. The results demonstrate that DATS can lower the amount of cross-shard transactions by 80% in a scenario involving 24 shards.