Occamy: Latency-Aware Scheduling with Controlled Degradation for Mixed-Criticality AI at the Edge
(2026) 11th ACM/IEEE Symposium on Edge Computing — Location: Santa Clara, CA, USA (13.October.2026)
Files
Occamy_CR_v2.pdf
Open Access - Adobe PDF
- 1.5 MB
Details
- Authors
- Abstract
- We study edge AI inference systems deployed across geographically distributed edge sites and serving latency-sensitive requests from spatially distributed users. Each application can feature multiple DNN model variants with different trade-offs among inference quality, inference time, and resource usage. When resources are insufficient, these variants enable graceful degradation of inference quality while maintaining strict adherence to tail latency Service Level Objectives (SLO). We present Occamy, a priority-aware scheduler that iteratively searches for a globally optimal deployment and introduces graceful degradation when necessary. For each application, Occamy jointly considers DNN model variant selection, placement across available edge sites, and scaling of application instances. This multi-dimensional optimization ensures continuity of service under P99 SLOs, even under resource pressure and scarcity. We evaluate Occamy on a real distributed edge testbed using Kubernetes, with realistic network latency and user request traces. Our results show that Occamy effectively preserves P99 latency guarantees for high-priority applications, applies minimal degradation to low-priority applications when necessary, and improves overall resource efficiency compared to state-of-the-art schedulers for edge workloads.
- Affiliations
Cnam Laboratoire CEDRIC
Citations
Cao, Y., Ntumba, P., Rivière, E., Sadre, R., & Secci, S. (2026, October 16). Occamy: Latency-Aware Scheduling with Controlled Degradation for Mixed-Criticality AI at the Edge. 11th ACM/IEEE Symposium on Edge Computing, Santa Clara, CA, USA. https://hdl.handle.net/2078.5/281337