Performance Parameters for Parallel Algorithms in GPU-Enhanced Environments

This work analyzes the role of graphic processing units (GPUs) in the framework of traditional parallel architectures (MIMD, SIMD, . . .) to find some effective parameters for the prediction of algorithms performance in a GPUenhanced environment. We consider a medical imaging application, namely the deconvolution of 3D Fluorescence Microscopy images, implemented in C with CUDA extension on a NVIDIA Tesla C1060: describing its design choices we intend to show how those parameters can affect actual algorithms performance