Evaluation of GPU Performance Compared to CPU for Implementing Algorithms with High Time Complexity

Abstract

Nowadays a number of applications with high volume of calculations are constantly increasing. Central Processing Unit (CPU) consists of finite number of cores. Degree of parallelism and implementation speed are issues that data high volume on CPU is low. Using the thread concept in programming, algorithms which have the parallelism capabilities, can be executed in parallel. There are many issues which in order to solving them, finding similar items in a metric space and grouping them in these issues is necessary. Computational complexity finding nearest neighbors is a challenge for run time. To evaluate the performance of GPUs speed in searching nearest neighbors, GPGPU and CUDA are used and compared with CPU usage. In this paper parallel implementation of the algorithm on GPU with access to its shared memory, is compared with parallel implementation of the algorithm on CPU through threads. It is understood that threads use graphics card's shared memory for communications, storing temporary data and retrieving data. Therefore, the parallelism on GPU is more useful than parallelism on CPU in High-Dimensional spaces. Finally, it is discussed that GPU reduces complexity to a considerable amount and is scalable.