Predicting Execution Time of CUDA Kernels with Unified Memory Capability

Abstract

Nowadays, GPUs are known as one of the most important, most remarkable, and perhaps most popular computing platforms. In recent years, GPUs have increasingly been considered as co-processors and accelerators. Along with growing technology, Graphics Processing Units (GPUs) with more advanced features and capabilities are manufactured and launched by the world's largest commercial companies. Unified memory is one of these new features introduced on the latest generations of Nvidia GPUs which allows programmers to write a program considering the uniform memory shared between CPU and GPU. This feature makes programming considerably easier. The present study introduces this new feature and its attributes. In addition, a model is proposed to predict the execution time of applications if using unified memory style programming based on the information of non-unified style implementation. The proposed model can predict the execution time of a kernel with an average accuracy of 87.60%.