Modeling Performance and Power on Disparate Platforms using Transfer Learning with Machine Learning Models

Paper available at the following link

Power has been an essential factor in addition to performance for the selection of computer systems today. Therefore, several research works use machine learning models to predict performance (runtime) and power using univariate or multivariate models. The cross-platform prediction is used to predict both targets for one platform or system with a trained model from the other. Transfer learning is used to train the machine learning model with one dataset to perform prediction on the other dataset with similar characteristics. In this work, we have used transfer learning to perform cross-system prediction from a simulation system to a physical system and cross-platform prediction from one physical system to the other. We have employed several machine learning univariate or multivariate models for our experiments. Our result shows that runtime and power prediction accuracy of more than 90% and 80% respectively is achieved for multivariate deep neural network model in cross-platform prediction. Similarly, for cross-system prediction runtime accuracy of 90% and power accuracy of 75% is achieved for the multivariate deep neural network. We have also analyzed prediction accuracy based on application type of workload, compute-bound versus memory-bound.

Citation

If you find this repo useful for your research, please consider citing our paper:

@INPROCEEDINGS{9198512,
  author={A. {Mankodi} and A. {Bhatt} and B. {Chaudhury} and R. {Kumar} and A. {Amrutiya}},
  booktitle={ In: Das B., Patgiri R., Bandyopadhyay S., Balas V.E. (eds) Modeling, Simulation and Optimization. Smart Innovation, Systems and Technologies, vol 206. Springer, Singapore.}, 
  title={Modeling Performance and Power on Disparate Platforms Using Transfer Learning with Machine Learning Models}, 
  year={2021},
  volume={206},
  number={231-246},
  pages={1-12},
  doi={https://doi.org/10.1007/978-981-15-9829-6_18}}

For any enquiries, please contact the main authors.