Support Vector Machines are a machine learning approach that is well studied, thoroughly vetted and effective in a large number of applications. The objective of this thesis is to accelerate an implementation of Support Vector Machines (SVM) using a heterogeneous computing system programmed using OpenCL in C/C++. LIBSVM, a widely-available, popular and open source implementation of SVM is chosen, allowing the presented work to be integrated seamlessly into existing systems. The proposed framework is evaluated in terms of speed and accuracy when performing training and classification on a number of standard data sets. Testing was based on two work station GPUs, the NVIDIA GTX 480 and Tesla K20, and a modern, work station CPU (Intel i5 Quad Core, 3 GHz).
We find that, for large data sets, training is accelerated by a factor ranging from 9 to 22. In general, speedup increases with the total number of training samples in the data set until the GPU device is fully utilized. While these gains in speedup are significant, they do not match the ideal parallel speedup, that is the total number of cores in the parallel system. Our findings indicate that performance is hampered by the portions of the SVM training algorithm that are sequential. In addition, we find that the classification phase of the SVM system is accelerated by a factor of up to 12. During classification only a relatively small number of samples are classified compared to the typical number of training samples, and the computational complexity of classification grows only linearly with the number of samples processed, as opposed to the training phase where it grows quadratically. The contri- butions of this thesis include the use of OpenCL for accelerating SVM training and testing on heterogeneous systems, and the performance analysis of the acceleration of SVM.
Library of Congress Subject Headings
Support vector machines--Testing; Machine learning; OpenCL (Computer program language)
Computer Engineering (MS)
Department, Program, or Center
Computer Engineering (KGCOE)
Sonia Lopez Alarcon
Peters, Ethan, "High Performance Implementation of Support Vector Machines Using OpenCL" (2014). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus