Towards Full-Stack Adaptivity in Byzantine Fault Tolerant Systems

Byzantine fault tolerant (BFT) systems, composed of distributed nodes, ensure security and reliability even under arbitrary node behaviors. Grounded in decades of foundational work, BFT systems are critical in today’s blockchains, cloud computing, cyber-physical systems, and more. However, as scale and user demands grow, the performance of traditional BFT systems is constantly questioned. Existing research efforts on improving performance primarily focus on designing static and siloed BFT systems tailored to a specific workload, fault scenario, and hardware setup. Unfortunately, prior studies have shown that such rigid and siloed systems are fragile in realistic deployments where workloads are dynamic, adversarial attacks evolve, and hardware configurations could change. The performance of a given system varies significantly as the environment changes, and the previously best-performing system could immediately become the worst. In other words, there is no one-size-fits-all BFT system that consistently delivers satisfactory performance. This lack of a clear “winner” among BFT systems, combined with their complexity, makes it difficult for application developers to select one and properly configure it. Worse yet, their choice might be invalidated as workloads or fault scenarios change.In this project, we aim to address this challenge by designing full-stack adaptive BFT systems that learn to adjust to environmental dynamics in real-time, in order to optimize runtime performance. Our plan includes two tasks: 1) Designing learned adaptive BFT consensus protocols. We will start by co-designing reinforcement learning algorithms with leader-based BFT consensus protocols, enabling adaptive protocol switching and parameter tuning while preserving safety and liveness. We will then extend our approach to directed acyclic graph (DAG)-based BFT protocols which exhibit more complexity and larger action space. To achieve the pinnacle of adaptivity, we will next investigate automated synthesis of new variants of leader-based and DAG-based BFT consensus protocols during runtime, where the new variants outperform existing protocols. This builds on our initial research on formalizing protocol design dimensions and valid design choices. 2) Developing an adaptive execution stack for BFT systems.The second half of our project goes beyond the consensus layer, aiming to adapt the entire execution stack to maximize end-to-end system performance. Building on insights from the first task, we will design learned adaptive transaction management paradigms for BFT systems. We will then develop an adaptive infrastructure layer for BFT systems based on our initial work on resource disaggregation. Finally, we plan to orchestrate cross-layer adaptation to unlock holistic system-level improvements.