Performance Prediction of Physical Computer Systems Using Simulation-Based Hardware Models

The advancements in computer systems with different hardware features have aided us with a selection from several options to execute a given software with a distinct performance profile on each system. However, it is a daunting task to acquire a large number of computer systems just for evaluating the performance of the software. Simulation-based hardware models can be built to predict the performance of physical systems, which we termed as “Cross Performance Prediction.” In this paper, we have shown that we can predict the performance of benchmarks programs from various application domains for physical systems using a machine learning model trained only on simulation-based hardware systems. We also have shown the categorization of benchmark programs using the correlation between hardware features and algorithm's runtime. Our cross performance prediction result achieves an average accuracy of 80%-90% for all benchmark programs.