Fast deployment and scoring of support vector machine models in CPU and GPU

Abstract

In this paper, we present an approach for the fast deployment and efficient scoring of Support Vector Machine (SVM) models. We developed a compiler for transforming a formal specification of a SVM and generating source code in different versions of the C/C++ language. This effectively automates the deployment of SVM models and its integration into the operational software for its use. The proposed compiler generates efficient code to deploy SVM models in CPUs (single or multi-core) and in Graphics Processing Units (GPUs) through NVIDIA's Computed Unified Device Architecture (CUDA). We also present an empirical evaluation of our compiler's targets scoring a SVM model with a linear kernel. In our experiments we score a real dataset in batch mode at different scales. The results show that our C CUDA implementation performs better as data scale increases and it is approximately 38 times faster than the single-core implementation using single precision floating-point values.