Main Content

MONARCH: Metric-driven Optimization for Network-Aware Raft Cluster Hierarchies

Author. Tobias Fischer

Supervisors: Prof. Dr.-Ing. Günter Schäfer, Prof.Dr. Boris Koldehofe

Format: Master's thesis, TU Ilmenau

Leader-based consensus protocols such as Raft are widely used in distributed systems, but their reliance on a single leader for coordination creates an inherent scalability bottleneck. As cluster sizes grow and network conditions become less uniform, the leader’s load can limit throughput. Existing approaches address individual aspects of this problem through dynamic parameter tuning, hierarchical node organization, or adaptive batching, but these mechanisms have rarely been evaluated in combination.

This thesis proposes MONARCH, a flexible, multi-layer Raft-based consensus framework that unifies three complementary strategies. A two-plane architecture separates liveness concerns – handled by a flat control plane – from log replication, which is handled by a hierarchical data plane. A QoS-aware partitioning algorithm constructs this hierarchy by greedily assigning nodes to subgroups based on measured per-link latency and packet loss. We introduce decentralized request handling, where clients may submit requests to multiple nodes in the hierarchy, distributing load across nodes and reducing commit latency. These mechanisms are complemented by adaptive tuning of election timeouts, heartbeat intervals, and batch sizes.

The framework is evaluated using a Go prototype on clusters of 7, 27, and 54 nodes under stable and adverse network conditions. Under stable conditions, the mechanisms integrate without an obvious performance penalty. Under adverse network conditions, batching is the most consistently effective mechanism, while the adaptive tuning mechanisms could not be reliably isolated from environmental variance. Decentralized request handling provides the clearest benefit: a QoS-aware hierarchy with batching sustains the full offered workload, while the classic Raft baseline falls well short. Synthetic micro-benchmarks confirm that the QoS-aware algorithm produces substantially higher-quality partitions than random assignment as the cluster grows.

The results demonstrate that hierarchical consensus can provide practical scalability benefits when the hierarchy is constructed using network-quality information.

Completed: 08.09.2026