EuroMLSys '21 | Predicting CPU usage for proactive autoscaling

[mental2008]

Presented in EuroMLSys '21. The 1st Workshop on Machine Learning and Systems (EuroMLSys), co-located with EuroSys '21.

Authors: Thomas Wang, Simone Ferlin, Marco Chiesa Ericsson AB, KTH Royal Institute of Technology

[mental2008]

A workshop paper.

Key Observation

Existing Kubernetes Vertical Pod Autoscaler (VPA) performs poorly on workloads that periodically change.

Solution

Propose a predictive method based on Long Short-Term Memory (LSTM).

• Avoid CPU insufficiency for various CPU workloads.
• Generate stabler predictions than the method based on Holt-Winters exponential smoothing (HW).

Some Details

• The algorithms run with the container’s CPU usage sampled at regular intervals.
• A new prediction is generated every 10 minutes.