Introduction
While Kubernetes has become the industry standard for orchestrating containerized workloads, it was originally designed for cloud environments where bandwidth is abundant and latency is negligible. In Collaborative Edge Computing (CEC)—where heterogeneous nodes like edge servers, gateways, and mobile devices share resources—this "cloud-first" logic falls short. Standard schedulers often ignore application-level performance, task dependencies, and the tight coupling between computation and networking resources at the edge.
System Architecture
To bridge these gaps, researchers developed ENTS (Edge-native Task Scheduling). ENTS is not a replacement for Kubernetes but an extension that introduces "edge-aware" components to the standard architecture.
The system incorporates a Data Flow Programming Model, treating applications as Directed Acyclic Graphs (DAGs) of functional modules. This makes internal dependencies explicit, allowing the scheduler to understand how data flows between tasks. A dedicated Profiler estimates execution times across heterogeneous hardware, while the core optimization engine handles task allocation and flow scheduling.
Joint Optimization of Compute and Network
The "Aha!" moment in ENTS is the realization that networking is a first-class citizen. Scheduling a task on a powerful node is ineffective if the network path to that node is a bottleneck. ENTS treats task allocation and flow scheduling as a single, joint optimization problem.
As shown in the figure above, simple job allocation without task partitioning (c) results in lower throughput. By introducing task partitioning and bandwidth allocation (e), and finally customized routing (f), the system can significantly increase throughput (from 2 to 4 in this example). ENTS uses Linux kernel functions like Traffic Control to actively "shape" the network to fit application needs.
Evaluation and Key Results
ENTS was evaluated using a real-world edge video analytics pipeline (object detection and recognition) on a testbed of Raspberry Pis, Nvidia Jetson Nanos, and Xavier NX nodes.
- Throughput: ENTS achieved 43% to 220% higher average job throughput than standard Kubernetes policies like Least Request Priority (LR) or Balanced Resource Allocation (BR).
- Waiting Time: In high-load scenarios with over 50 concurrent jobs, ENTS significantly reduced waiting times by avoiding resource fragmentation.
- Scalability: The system maintained superior performance as the network scaled from 10 to 70 edge nodes.
Limitations and Future Work
While ENTS represents a significant advancement, future work will focus on increasing resource granularity (optimizing specific CPU/memory allocations), deeper integration with Software-Defined Networking (SDN), and developing robust online profiling tools to estimate workloads in real-time.
Summary updated