Frequently Asked Questions¶
General Questions¶
What is ICOS-FL?¶
ICOS-FL is a federated learning framework built on Flower for resource monitoring and prediction across distributed nodes. It uses LSTM models to predict CPU, memory, and power usage patterns.
How does ICOS-FL compare to other federated learning frameworks?¶
ICOS-FL is specifically optimized for time series prediction of system metrics. While general-purpose frameworks like TensorFlow Federated or PySyft offer broader ML capabilities, ICOS-FL provides a streamlined solution for resource monitoring with built-in DataClay integration.
Does ICOS-FL require specialized hardware?¶
No. ICOS-FL runs on standard hardware but can utilize GPUs if available for faster training. All components can run on CPU-only machines with reasonable performance.
Technical Questions¶
How is data stored in ICOS-FL?¶
ICOS-FL uses DataClay as its persistent storage layer. Time series data is stored in a sliding window format, with the TimeSeriesData class handling automatic management of the buffer size.
What metrics does ICOS-FL collect?¶
By default, ICOS-FL collects CPU usage, memory usage, and power consumption through Scaphandre and OpenTelemetry. You can customize the metrics collection by modifying the BridgeConfiguration.
Can I deploy ICOS-FL without Docker?¶
Yes, all components can be installed and run natively. However, Docker simplifies deployment and ensures consistent behavior across different environments.
Federated Learning Questions¶
How many clients does ICOS-FL support?¶
ICOS-FL has been tested with up to 20 client nodes. The practical limit depends on network bandwidth, server resources, and aggregation strategy.
How does ICOS-FL handle client failures?¶
ICOS-FL uses Flower’s built-in fault tolerance mechanisms. If a client becomes unavailable during training, the server can continue with the remaining clients as long as the minimum number of available clients is satisfied.
Can I use my own custom models?¶
Yes, ICOS-FL supports custom models as long as they implement the required interface. See Custom Models for details.
Can ICOS-FL work with non-time-series data?¶
While ICOS-FL is optimized for time series forecasting, its architecture can be adapted for other federated learning tasks. However, you would need to modify the data processing pipeline.
Performance and Scaling¶
How much data does ICOS-FL need for accurate predictions?¶
For reasonable forecasting accuracy, ICOS-FL typically requires at least a few hours of historical data. Prediction accuracy improves with more data and training rounds.
What is the network overhead of federated learning?¶
The network overhead is proportional to model size rather than data size. LSTM models in ICOS-FL are relatively compact (typically <1MB), making the communication efficient.
How far ahead can ICOS-FL predict resource usage?¶
By default, ICOS-FL is configured to predict 5 minutes ahead based on 15 minutes of historical data. This prediction horizon can be adjusted by modifying the time window configuration.