Cerberus: Latency-Aware Scheduling with Graceful Degradation for Constrained Edge Infrastructures
(2026) IEEE 45th International Symposium on Reliable Distributed Systems (SRDS) — Location: Rome, Italy (22.September.2026)
Files
Cerberus_CR_v4.pdf
Open Access - Adobe PDF
- 1.21 MB
Details
- Authors
- Abstract
- Resource scarcity is a critical challenge for workload scheduling in edge clusters, where contention among applications can lead to severe performance degradation and uncontrolled Service-Level Objectives (SLO) violations. We present Cerberus, a scheduler that integrates accurate queuing-based latency modeling with globally optimized scheduling decisions. Cerberus predicts the relationship between resource allocation and tail latency using a precise model of Kubernetes load-balancing mechanisms, and computes placement plans that satisfy tail-latency SLOs for multiple applications while minimizing resource consumption. When resources are insufficient to simultaneously meet all SLOs, Cerberus introduces a novel graceful degradation mechanism that prioritizes workloads by criticality and applies controlled, bounded degradations to low-criticality services instead of evicting them. Our evaluation on a Kubernetes-based edge testbed demonstrates that Cerberus improves fairness, system stability, and resource efficiency compared to eviction-based and resource-agnostic scheduling strategies.
- Affiliations
Citations
Cao, Y., Riviere, E., & Sadre, R. (2026, September 25). Cerberus: Latency-Aware Scheduling with Graceful Degradation for Constrained Edge Infrastructures. IEEE 45th International Symposium on Reliable Distributed Systems (SRDS), Rome, Italy.