Towards High-Throughput and Low-Latency Heterogeneous Blockchains via RL-Driven Network Sharding

Sharded blockchains enable scalable distributed ledgers by partitioning transaction workloads into multiple parallel processing shards. However, existing sharding protocols for heterogeneous blockchain environments mainly focus on mitigating performance degradation caused by uneven node bandwidth, while overlooking dynamic network variations, such as bandwidth fluctuations, unstable packet delivery, and transaction queue accumulation. These factors result in inefficient shard utilization and significant performance disparities among shards. To address these issues, we propose PabShard, an intelligent sharding framework for heterogeneous blockchain systems. PabShard leverages reinforcement learning to adaptively regulate shard operations according to network states and workload conditions. Specifically, we address two key challenges. First, to achieve effective shard organization under heterogeneous network conditions, we design a PPO-driven shard control mechanism that dynamically adjusts shard configurations based on observed system states. Meanwhile, an asymmetric bandwidth-adaptive asynchronous Merkle tree (BA-AMT) is introduced to improve transaction verification efficiency by exploiting available bandwidth resources. Second, to alleviate transaction congestion caused by heterogeneous shard processing capacities, we develop an adaptive block generation rate control mechanism that dynamically regulates the processing capability of individual shards. We implement PabShard on a Huawei Cloud cluster with heterogeneous bandwidth configurations ranging from 20 to 60 Mbit/s, where each physical host supports numerous virtualized blockchain nodes. Extensive experiments against representative sharding protocols, including Monoxide, GearBox, and Manifoldchain, demonstrate the effectiveness of PabShard. Under a transaction arrival rate of 3000 TPS with eight shards, PabShard achieves a 12.3% throughput improvement and reduces confirmation latency by 14.8% compared with the best performing baseline.