Interpreting Write Performance of Supercomputer I/O Systems with Regression Models

This work seeks to advance the state of the art in HPC I/O performance analysis and interpretation. In particular,we demonstrate effective techniques to: (1) model output per-formance in the presence of I/O interference from productionloads; (2) build features from write patterns and key parametersof the system architecture and configurations; (3) employ suitablemachine learning algorithms to improve model accuracy. We train models with five popular regression algorithms and conduct experiments on two distinct production HPC platforms. We find that the lasso and random forest models predict outputperformance with high accuracy on both of the target systems.We also explore use of the models to guide adaptation in I/Omiddleware systems, and show potential for improvements of atleast 15% from model-guided adaptation on 70% of samples,and improvements up to 10x on some samples for both of the target systems.

Full Paper