Parallel Self-organizing Map Using Shared Virtual Memory Buffers

TL;DR: The experimental results show the parallel SOM running on heterogeneous platform has significant improvement in computation time, and the proposed parallel SOM architecture is suitable for both GPU and heterogeneous system with the aim of comparing the performance in term of computation time.

Abstract: Parallel implementation of Self-organizing Map (SOM) has been studied since last decade. Graphic Processing Unit (GPU) is one of most promising architecture for executing SOM in parallel. However, there are performances issues are highlighted when imposing larger mapping and dataset size onto parallel SOM that executed on the GPU. Alternatively, heterogeneous systems that soldered GPU together with Central Processing Unit (CPU) are introduced in order to improve communication between CPU and GPU. Shared Virtual Memory (SVM) is one of features in OpenCL 2.0 which allows the host and the device to share a common virtual address range. Thus this research proposes to introduce a parallel SOM architecture that suitable for both GPU and heterogeneous system with the aim to compare the performance in term of computation time. The architecture comprises of three kernels that executed on two different platforms (1) discrete GPU platform and (2) heterogeneous system platform that tested using SVM buffers. The experimental results show the parallel SOM running on heterogeneous platform has significant improvement in computation time.