RAINDROP aims to address in a unique way the challenge of building, deploying, and operating applications on upcoming cloud infrastructures formed of a continuum of core (data centers) and edge resources.
RAINDROP aims to address in a unique way the challenge of building, deploying, and operating applications on upcoming cloud infrastructures formed of a continuum of core (data centers) and edge resources.
Our paper “Occamy: Latency-Aware Scheduling with Controlled Degradation for Mixed-Criticality AI at the Edge”, in collaboration with Patient, has been accepted as a full paper at SEC 2026.
We extend the graceful-degradation methodology of Cerberus from latency relaxation to model-variant selection through the development of the Occamy global scheduling model, targeting mixed-criticality AI inference across distributed edge sites. Whereas Cerberus restores feasibility by relaxing the latency bounds of low-criticality applications, Occamy exploits the multiple model variants that AI applications expose, each trading inference accuracy for lower latency and resource consumption. Occamy predicts the P99 end-to-end latency of a candidate deployment and jointly selects a model variant, an instance count, and a user-to-site assignment for each application under strict per-application SLOs and edge capacity limits. Under resource scarcity, a priority-aware degradation algorithm switches low-criticality applications to lighter variants in controlled steps until a feasible placement is found, while preserving the nominal quality of high-criticality ones. This delivers graceful degradation along the accuracy dimension, keeping all applications operational with quality loss that is predictable and minimal.
Yinan will present the paper during the conference on October 13-16 2026 in Santa Clara, CA, United States.