SW Runtime Estimation using Automata Theory and Deep Learning on HPC

People use high performance computers (HPCs) for computation-intensive tasks. Often these tasks require a lot of running time of their corresponding softwares. It is important to execute several tasks simultaneously for the system utilization while finishing all tasks within their desired deadlines. Therefore, it is important to know the runtime of each computation-intensive task without executing them in order to schedule the tasks on HPC and obtain better system performance. We propose a method for predicting runtime of MPI-based softwares on HPC using automata theory and deep learning. We first analyze a source code of an input program by representing the code as finite automata and measuring their state complexities. Next, we train the execution runtime of each module of our finite automata using deep neural network (DNN) together with its own state complexity. Then we combine all modules and make a single SW-runtime-prediction model. For experiment, we train the proposed model using OSU benchmark data, HPL and two in-house datasets, and present the usefulness of our model. We also demonstrate the adaptability of our model by updating the current model for new inputs using incremental DNN, which is an important feature for coping with new softwares or new systems.