ShardOpt: Optimal Sharding Strategy with Transaction Prediction

To address the scalability challenges of blockchain technology, sharding has emerged as a solution. Most existing sharding methods focus on transaction-based sharding, where transactions are distributed across various shards, with each shard responsible for processing a subset of transactions. However, the lack of prediction regarding transaction volumes among shard nodes leads to slow transaction speeds and reduced throughput, especially when managing extensive cross-shard transactions. In this paper, we propose a predictive model for node transaction volumes utilizing the Gate Recurrent Unit (GRU) recurrent neural network. We employ this model to optimize the sharding process of the blockchain. Specifically, we develop a sophisticated 5-layer GRU network, using actual transaction volumes as training data and iteratively refining the model to improve its predictive accuracy. Based on the model’s predictions, we partition blockchain nodes into shards, grouping nodes with similar transaction volumes within each shard while ensuring a decreasing trend in total transaction volumes within shards. Finally, we evaluate system throughput using real-time monitoring techniques. Experimental results show a significant throughput enhancement of up to 83% for our designed sharded blockchain system.