Prediction of the Resource Consumption of Distributed Deep Learning Systems

Abstract

Predicting resource consumption for the distributed training of deep learning models is of paramount importance, as it can inform a priori users of how long their training would take and enable users to manage the cost of training. Yet, no such prediction is available for users because the resource consumption itself varies significantly according to "settings"such as GPU types and also by "workloads"like deep learning models. Previous studies have attempted to derive or model such a prediction, but they fall short of accommodating the various combinations of settings and workloads together. This study presents Driple, which designs graph neural networks to predict the resource consumption of diverse workloads. Driple also designs transfer learning to extend the graph neural networks to adapt to differences in settings. The evaluation results show that Driple effectively predicts a wide range of workloads and settings. In addition, Driple can efficiently reduce the time required to tailor the prediction for different settings by up to 7.3×.