Dataset Origin

This dataset is derived from OpenCellID-grounded topologies, which are constructed using cellular infrastructure data from OpenCellID: https://opencellid.org/#zoom=16&lat=37.77889&lon=-122.41942

Overview

This repository provides a topology-and-workload dataset designed for evaluating distributed orchestration systems for multimodal generative models. The dataset comprises four representative server-topology configurations that emulate real-world distributed compute infrastructures with varying geographical characteristics, network scales, and resource heterogeneity. It is specifically intended to support research on reinforcement-learning (RL)-based resource optimization, scheduling, and orchestration in distributed multimodal model serving environments.

Dataset Description

1. Server Topologies

The dataset includes four primary server-topology configurations (S1–S4), each representing a distinct deployment scale and scenario:

S1 — Switzerland Region (Small-Scale, High-Density)

• Number of nodes: 25 distributed server nodes
• Compute capacity: 16–54 units (mean: 34.6 units)
• Storage capacity: 16–34 units (mean: 25.3 units)
• Network connectivity: sparse graph with an average degree of 3.2
• Deployment type: regional datacenter-style deployment
• Intended use: evaluation under small-scale, high-density deployments

S2 — United Kingdom Region (Medium-Scale, Geographically Distributed)

• Number of nodes: 51 distributed server nodes
• Compute capacity: 16–53 units (mean: 35.8 units)
• Storage capacity: 15–38 units (mean: 26.2 units)
• Deployment type: nation-scale infrastructure
• Intended use: evaluation under medium-scale geographically distributed deployments

S3 — Germany Region (Large-Scale, Heterogeneous)

• Number of nodes: 101 distributed server nodes
• Compute capacity: 19–55 units (mean: 36.7 units)
• Storage capacity: 12–39 units (mean: 26.0 units)
• Deployment type: large-scale heterogeneous-resource environment
• Intended use: validation of orchestration complexity under large-scale deployments

S4 — Milan Metropolitan Network (Urban Edge Computing)

• Number of nodes: 31 densely deployed servers
• Deployment type: urban edge-computing scenario
• Intended use: evaluation of edge intelligence in smart-city settings

2. Model Deployment

Each topology is accompanied by a deployment of generative models distributed across nodes to emulate heterogeneous real-world placements:

• Total models per topology: 227 generative models
• Text-to-Text models: 191
• Text-to-Image models: 30
• Image-to-Video models: 6

Note: To reflect practical heterogeneity, the model placement strategy differs across topologies.

3. Task Dataset and Train/Test Splits

Each topology includes user-task traces for training and evaluation:

• Training set: 1,000 user tasks per topology
• Test set: 200 independent tasks per topology
• Special setting: S4 is additionally designated as a test-only topology for cross-domain generalization evaluation.

---

Contact & Support

For questions, bug reports, or collaboration inquiries:

• Email: [gymorsiback@tju.edu.cn]

---