paper.txt

Abstract:



Introduction:
	MPI, CUDA, OpenCL, and OpenMP are key technologies in parallel processing and High Performance Computing (HPC). Comparisons between these technologies is, in many ways, a difficult process, since there is specialized knowledge bases for each technology. Further, these technologies represent different paradigms in the way parallel processing is done. 
     	At least MPI, OpenCL, and OpenMP share the goal of being massively cross-platform. CUDA aims to be a tool specialized to run on certain (NVidia) hardware.

CUDA versus OpenCL
     CUDA, while not easy to install, had very clear instructions. All the components snapped in nicely to my Ubuntu installation (even if they aren't available on apt), even giving my desktop a nice bump in graphics sparkle.

     CUDA is not a piece of cake to use at first. It requires that the programmer wrap his brain around the notion of a kernel. A kernel is simply device code that is run out to the GPU, not run locally on the CPU hardware. 

     OpenCL was not a piece of cake to install. I should have recognized the need to install a specific set of drivers, but I assumed that NVidia had given me the newest drivers with the CUDA installation. From some comments on NVidia forums, it appears that even bleeding-edge NVidia drivers with more recent release numbers don't contain the libraries necessary for OpenCL programs to compile correctly.