A Comprehensive Hyperledger Fabric Performance Evaluation based on Resources Capacity Planning
Carlos Melo, Glauber Gonçalves, Francisco A. Silva, André Soares
TL;DR
This work addresses the challenge of performance tuning for Hyperledger Fabric in permissioned networks by developing a stochastic Petri Net (SPN) framework that models the endorsement–ordering–commit workflow under varying block size, block timeout, and resource capacity. It combines the SPN model with a $2^3$ factorial Design of Experiments to quantify how these factors and their interactions affect mean response time ($MRT$) and other metrics, using Little's Law to relate throughput, queue lengths, and arrival rate. Key contributions include a holistic SPN representation of the full transaction flow, identification of bottlenecks at endorsement, ordering, and commit stages, a calibration strategy for block size and timeout, and a sensitivity analysis that highlights timeout as the dominant factor influencing performance. The results offer practical guidance for pre-deployment capacity planning and configuration tuning in HL Fabric deployments, enabling administrators to anticipate latency and throughput outcomes before implementation.
Abstract
Hyperledger Fabric is a platform for permissioned blockchain networks that enables secure and auditable distributed data storage for enterprise applications. There is a growing interest in applications based on this platform, but its use requires the configuration of different blockchain parameters. Various configurations impact the system's non-functional qualities, especially performance and cost. In this article, we propose a Stochastic Petri Net to model the performance of the Hyperledger Fabric platform with different blockchain parameters, computer capacity, and transaction rates. We also present a set of case studies to demonstrate the feasibility of the proposed model. This model serves as a practical guide to help administrators of permissioned blockchain networks find the best performance for their applications. The proposed model allowed us to identify the block size that leads to a high mean response time (ranging from 1 to 25 seconds) caused by a change in the arrival rate.