Model Info This repository contains a collection of machine learning models developed in collaboration between University College Dublin (UCD) and Chocolate Cloud (CC). These models are deployed within the SkyFlok Gateway component (hosted in London). They perform Latency Prediction to estimate the time required to retrieve a file from specific cloud storage backends. By predicting download times based on temporal patterns and file size, these models enable the Gateway to intelligently route download requests to the fastest available storage region. This minimizes retrieval latency and optimizes network efficiency for the end user. The models are exported in ONNX format (Opset 18) to ensure high-performance, low-latency inference within the real-time routing logic. Purpose These models perform Regression to predict a continuous value: Input: File size and detailed timestamps (Hour, Day, Time of Day). Output: Estimated Transfer Time (Latency) in milliseconds. Repository Structure & Backend Mapping Since latency characteristics vary between cloud providers, a separate model is trained for each storage backend. Use the table below to identify which model corresponds to which cloud provider/region. Backend ID Model Filename Cloud Provider Region Location 4 model_backend_id_4.onnx Google Cloud europe-west1 St. Ghislain, Belgium 🇧🇪 20 model_backend_id_20.onnx AWS eu-west-1 Dublin, Ireland 🇮🇪 39 model_backend_id_39.onnx Microsoft Azure WEST EUROPE Amsterdam, Netherlands 🇳🇱 79 model_backend_id_79.onnx OVH Cloud GRA Gravelines, France 🇫🇷 137 model_backend_id_137.onnx Exoscale Geneva Geneva, Switzerland 🇨🇭 144 model_backend_id_144.onnx Scaleway Warsaw Warsaw, Poland 🇵🇱 Directory Layout /models ├── model_backend_id_4.onnx # Model for Backend ID 4 ├── model_backend_id_20.onnx # Model for Backend ID 20 ├── model_backend_id_39.onnx # ... ├── model_backend_id_79.onnx ├── model_backend_id_137.onnx └── model_backend_id_144.onnx └── model_config.json # Hyperparameters & metadata Training Data The models were trained on historical transfer logs collected from the SkyFlok platform. The dataset captures real-world network performance metrics across different times of day and days of the week. Features Used: Workload: File size (bytes). Temporal: Time of day (categorical: morning, afternoon, etc.), Hour, Minute, Second, Day of Week. Model Architecture These models utilize a Scikit-Learn Pipeline architecture, fully converted to ONNX: 1. Preprocessing: A ColumnTransformer that handles mixed data types (One-Hot Encoding for strings, pass-through for numbers). 2. Regressor: Gradient Boosting Regressor (2000 trees, Max Depth 12). This architecture allows the model to capture complex, non-linear relationships between network congestion (time of day) and transfer speeds. Model Specification (Common Interface) All models in this collection share the same input/output interface specifications. Inputs The models accept a dictionary of standard Python lists. Note: The double brackets [[ ]] are required to represent a batch size of 1. Input Name Type Shape Description Example time_of_day String [batch, 1] Categorical time block "morning", "night" hour Int64 [batch, 1] Hour of the day (0-23) 14 minute Int64 [batch, 1] Minute of the hour 30 second Int64 [batch, 1] Second of the minute 45 day_of_week String [batch, 1] Full day name "Monday", "Friday" size Float32 [batch, 1] File size in Bytes 100000.0 Outputs Name: predicted_latency Data Type: float32 Shape: [batch_size, 1] Unit: Milliseconds (ms) Limitations Backend Specificity: model_backend_id_4.onnx is trained specifically on the performance history of Backend #4. It should not be used to predict latency for other backends. Historical Bias: Predictions are based on historical trends; sudden network outages or unprecedented congestion events may result in prediction errors. Usage Demo To run any model from this collection, follow the steps below. 1. Setup Environment python3.13 -m venv venv source venv/bin/activate pip install onnxruntime 2. Run Inference python inference_demo.py Code Example If you prefer to integrate it into your own application, here is the minimal code required: import onnxruntime as ort from datetime import datetime # --- Configuration --- backend_id = 4 file_size = 100000.0 # 100 KB # --- Helper --- def get_time_of_day(hour): if 5 <= hour < 12: return 'morning' elif 12 <= hour < 17: return 'afternoon' elif 17 <= hour < 21: return 'evening' return 'night' # --- 1. Prepare Input --- # Note: Double brackets [[ ]] create the required batch dimension (Batch=1) t = datetime.now() inputs = { 'time_of_day': [[get_time_of_day(t.hour)]], 'hour': [[t.hour]], 'minute': [[t.minute]], 'second': [[t.second]], 'day_of_week': [[t.strftime('%A')]], 'size': [[file_size]] } # --- 2. Run Inference --- session = ort.InferenceSession(f"models/model_backend_id_{backend_id}.onnx") result = session.run(None, inputs) print(f"Predicted Latency: {result[0][0][0]:.2f} ms") Citation If you wish to cite this specific model collection, please use the citation generated by Zenodo (located in the right sidebar of this record). Acknowledgement & Funding This work is part of the MLSysOps project, funded by the European Union’s Horizon Europe research and innovation programme under grant agreement No. 101092912. More information about the project is available at https://mlsysops.eu/