Enhancing Performance in Heterogeneous Computing: A Comparative Study of CUDA on GPUs and CPUs

Heterogeneous computing, in which Graphics Processing Units (GPUs) serve as coprocessors to CPUs for arithmetic-intensive data-parallel tasks, has become increasingly popular due to the pressing demand for higher computational efficiency in science and engineering. When it comes to heterogeneous computing on NVIDIA GPUs, the emerging standard is CUDA, or Compute Unified Device Architecture. The primary objective of CUDA is to free developers from the burden of implementing parallel algorithms and instead allow them to concentrate on the work at hand. The main objective of the work is to obtain the performance gain in execution speed for the dynamic algorithms which generally are complex and takes a very long time for execution and compare results on different GPU processors and CPU and have a comparative study of algorithms. This work is mainly concentrated on working with dynamic algorithms such as Knapsack, Binary Search, Longest Common Subsequence, Kruskal's Algorithm on GTX 480 and Tesla C2070 GPUs. Also, the work consists of preliminary sorting algorithms such as Insertion Sort, Selection Sort, and Bubble Sort. Report contains study of the NVIDIA CUDA architecture, CUDA SDK tool kit and different methods to get performance benefit, implementation of intermediate tool to find out functions and their dependencies from the source code and the implementation of the complete algorithm, testing and obtaining statistics for the same. The performance of such algorithms was improved by a factor of 2X to 64X when implemented on GPU as opposed to CPU.