A Queueing‑Driven Controller for Mempool, Block Fullness and Tail Latency

Abstract

Permissioned blockchain platforms with Byzantine-fault-tolerant consensus increasingly host latency-sensitive workloads. Yet their performance is governed by coupled, burst-prone subsystems: the transaction mempool, batch (block) formation and leader-based finality. We design, model and evaluate a queueing-driven controller that uses latency quantiles and block utilisation as service-level indicators to steer both mempool policy and block-formation cadence. The modelling view abstracts the mempool and block builder as a finite-buffer batch-service queue calibrated from node telemetry. At the same time, the controller manipulates safe levers — time-bounded adjustments of the block period and conservative changes to admission/eviction in the mempool — guarded by error-budget logic and consensus constraints. On a Hyperledger Besu IBFT/QBFT testbed we observe that, under trace-driven bursts, p99 time-to-inclusion halves from 3.8 s to 1.9 s (-50%) and p95 drops from 2.4 s to 1.6 s (-33%); average block utilisation rises from 82% to 92% while its coefficient of variation halves; sustainable throughput at ρ≈0.9 improves from 270 to 305 TPS without breaching a 2.5 s p99 SLO; mempool drain-time after a burst shrinks from 18 s to 7 s; and IBFT/QBFT round-changes fall from 38 to 11 per 10k blocks. These gains are achieved with ≈1–2% CPU overhead and negligible extra RPC traffic. The results align with queueing-theoretic expectations for finite-buffer batch service and demonstrate a practical path to SLO-centred control in permissioned chains.