Modern edge computing faces a silent efficiency killer: resource waste. Traditional orchestration frameworks like Kubernetes often rely on conservative, rule-based scheduling where applications overestimate their CPU and RAM needs. This leads to significant “slack resources”—allocated capacity that sits idle while other tasks wait in queues. The Horizon Europe project ACES is tackling this challenge through a novel, self-organizing local decision mechanism inspired by the natural world.
A multi-agent approach to heterogeneous workloads
By modeling edge infrastructure as a multi-agent system, ACES moves away from rigid, top-down management. The system distinguishes between rigid pods, which have strict execution windows and resource requirements, and elastic pods, which are flexible enough to exploit residual resources. At the heart of this architecture is the cluster master agent and specialized worker agents that manage resources collectively. Instead of a static queue, the system operates as a dynamic environment where agents interact to find the most efficient path for deployment.
Swarm intelligence and the ABC algorithm
The breakthrough in the ACES scheduler lies in a swarm intelligence-inspired algorithm based on the Artificial Bee Colony (ABC) meta-heuristic. In this model, rigid pods act like bees that monitor their own “food quality”—in this case, their available resource slack. The worker agent maintains a compact lookup table, grouping these pods into temporal “buckets” based on their small or large resource footprints. When an elastic pod arrives, the worker doesn’t need to perform an exhaustive, slow search. Instead, it uses probabilistic selection to find a compatible rigid pod “host,” allowing the elastic task to piggyback on the unused slack.
To ensure that latency-critical tasks are never compromised, the ACES master agent employs a sophisticated prioritization logic. It first satisfies all possible rigid pod demands before turning to the elastic queue. A controlled probability mechanism allows the system to occasionally treat elastic pods as rigid ones, balancing satisfaction rates across the entire swarm. This bottom-up resource allocation strategy, combining vertical autoscaling with decentralized coordination, ensures that even the most dynamic edge environments can maximize their hardware potential without sacrificing Quality of Service.
Project deliverables
More information can be found in two deliverables from the project website:
- Deliverable 4.3 “ACES – Coordination methodology and library implemented as software for realizing autonomy and actionability”
- Deliverable 4.5 “ACES – Distributed knowledge base and data management systems”
Publication
A. Ghasemi and M. Schranz: Bottom-Up Resource Orchestration in Edge Computing: An Agent-Based Modeling Approach. In Proc. IEEE International Conference on Intelligent Systems, Varna, Bulgaria, 2024.
Feature image created with GenAI