Self-tuning Filers - Overload Prediction and Preventive Tuning Using Pruned Random Forest

摘要

The holy-grail of large complex storage systems in enterprises today is for these systems to be self-governing. We propose a self-tuning scheme for large storage filers, on which very little work has been done in the past. Our system uses the performance counters generated by a filer to assess its health in real-time and modify the workload and/or tune the system parameters for optimizing the operational metrics. We use a Pruned Random Forest based solution to predict overload in realtime - the model is run on every snapshot of counter values. Large number of trees in a random forest model has an immediate adverse effect on the time to take a decision. A large random forest is therefore not viable in a real-time scenario. Our solution uses a pruned random forest that performs as well as the original forest. A saliency analysis is carried out to identify components of the system that require tuning in case an overload situation is predicted. This allows us to initiate some 'action' on the bottleneck components. The 'action' we have explored in our experiments is 'throttling' the bottleneck component to prevent overload situations.