Learned Sharding Toward Sustainable Communications and Networking in Blockchains

Abstract

Sharding scales blockchain by grouping blockchain nodes into committees each of which processes a portion of the total transactions in parallel. The issue with sharding is the enormous volume of cross-shard transactions, which results in high communication costs to ensure transaction atomicity. The account-based sharding problem can be viewed as the vertex classification problem of the account-transaction graph. However, prior studies employed traditional graph partitioning algorithms for sharding, failing to make full use of the account relationship in the graph structure. In this work, we aim to address the sharding problem from the perspective of deep learning that can learn the graph structure toward sustainable communications. We propose an efficient deep learning-based sharding scheme (DLS) based on the graph attention (GAT) network. The account and transaction information are input into the GAT for semi-supervised training and account/vertex classification. Since the performance may degrade in the case of limited label information, we incorporate the label propagation method to acquire the label information of non-trained accounts. We also extend our approach to deal with the new account scenario without retraining the neural network. Extensive experiments on Ethereum data demonstrate that our proposed DLS can effectively reduce the number of cross-shard transactions.