Prediction Model of an HPC Application on CPU-GPU Cluster using Machine Learning Techniques

摘要

In today's world hybrid computing cluster, is comprised of high-intensity computation central processing unit (CPU) and graphical processing unit (GPU) based nodes. In this article, a novel analytical prediction model by considering parameters such as the number of CPUs+GPUs cores, peripheral component interconnects express(PCI-E) bandwidth and CPU-GPU memory access bandwidth for varying data input sizes and recorded as historical data. A partial amount of data is tested to train our novel prediction model. Predicted execution time against actual execution time has been compared to enhance the accuracy of the model and reduce or remove any errors. The proposed prediction model is a major module that has been utilized from scheduling the strategy to scheduling the high-performance computing (HPC) application, which gives the least predicted execution time on the best resources of a heterogeneous cluster. The proposed predictive scheduling scheme with scheduling strategy has been tested by using the game of life benchmark applications on a CPU-GPUs cluster. The prediction model has been compared against machine learning techniques and it is observed that the proposed novel analytical prediction model has achieved less than 19% prediction error. The performance of our predictive scheduling scheme with other best existing schemes TORQUE has been compared, and it is also observed that the predictive scheduling scheme is 63% more efficient than the TORQUE.