EDGELESS KPI-3 experiments

Dataset Description

This dataset contains the experimental results used to evaluate centralized and decentralized orchestration strategies in an EDGELESS cluster. The experiments compare a baseline configuration with a single orchestration domain against a target decentralized deployment with multiple orchestration domains.

Experimental Scenarios

Two system configurations were evaluated:

• Baseline (Single-Domain):
  An EDGELESS cluster with a single orchestration domain. All scheduling and resource management decisions are taken by a single ε-ORC instance. In this configuration, the ε-CON has no decision-making autonomy.

• Decentralized (Multi-Domain):
  An EDGELESS cluster composed of six orchestration domains. Resource management decisions are performed at two levels:

  • The ε-ORC manages local resources within each domain using fine-grained measurements.

  • The ε-CON coordinates function instances across domains using aggregated measurements and operates at a longer time scale.

All other experimental conditions (benchmarking methodology, workload, applications, node resources, and runtime environments) were kept identical across scenarios.

Application Workflow

The evaluated application consists of a serverless workflow processing image streams:

1. file-pusher (resource):
   Generates one base64-encoded image every 100 ms (10 images/s) from a local dataset of road images.

2. flow-control (WebAssembly, stateless):
   Limits the number of in-flight images to 5, with a reset timeout after 50 input images. Excess images are dropped.

3. image-scale (WebAssembly, stateless):
   Rescales images to a maximum resolution of 676×380 pixels while preserving the aspect ratio.

4. obj-detect (resource):
   Performs object detection using YOLO (Ultralytics container optimized for NVIDIA AGX Orin). Detected object metadata (position, size, type) are added to each message.

5. obj_track (WebAssembly, stateful):
   Tracks objects across the previous 5 frames, adding bounding boxes and trajectories.

6. http-poster (resource):
   Decodes images and sends them to an external HTTP service.

7. Viewer (external service):
   A Rust-based web application that logs received images with timestamps.

Workload Configuration

The workload was generated using the edgeless_benchmark utility with the following parameters:

• Experiment duration: 1800 seconds.

• Workflow duration: Poisson-distributed with mean 60 s, 120 s, or 240 s.

• Workflow interarrival time: Poisson-distributed with mean 2 s.

This configuration results in approximately 30, 60, or 120 concurrently active applications.

Testbed

Experiments were conducted at the CNR-IIT facilities (Ubiquitous Internet research group).

• Control Plane:

  • ε-CON deployed in a dedicated container within a Proxmox cluster (not restarted across experiments).

  • ε-ORC services deployed in dedicated containers with Redis proxy and dataset dumping enabled.

  • edgeless_benchmark executed in the same Proxmox environment.

• Compute Nodes:

  • Raspberry Pi 5.

  • NVIDIA AGX Orin 64 GB.

  • Runtime: RUST_WASM.

  • Resource providers: file-pusher and http-poster.

  • obj-detect deployed only on Orin devices.

  • Refresh interval to ε-ORC: 10 s.

  • Performance sampling enabled.

• Networking:

  • Cisco Layer-2 switches.

  • 1 GbE links for Raspberry Pi 5 nodes.

  • 10 GbE links for AGX Orin nodes.

• Power Monitoring:

  • Active power consumption measured via Raritan PDUs with per-outlet monitoring.

Dataset Contents

The dataset includes:

• Raw experimental logs collected from orchestration services and nodes.

• Performance samples and resource utilization metrics.

• Power consumption measurements.

• Data required to reproduce the plots presented in the associated publication.

Plots can be reproduced using the analyze.py script available in the plots/ directory of the corresponding GitHub repository.

Acknowledgments

The authors acknowledge the technical support of the Computer and Communication Networks Technology Unit at CNR-IIT for managing the Proxmox cluster infrastructure and for setting up the network connectivity of the Raspberry Pi and NVIDIA AGX Orin devices used in the experiments.