Exploring the problem of GPU programming for data-intensive applications: A case study of multiple expectation maximization for motif elicitation

Recently General-Purpose Computing on Graphics Processing Units (GPGPU) has been used to reduce the processing time of various applications, but the degree of acceleration by the Graphical Processing Unit (GPU) depends on the application. This study focuses on data analysis as an application example of GPGPU, specifically, the design and implementation of GPGPU computation libraries for data-intensive workloads. The effects of efficient memory allocation and high-speed read-only memories on the execution time are evaluated. In addition to employing a single GPU, the scalability using multiple GPUs is also evaluated. Compared to a Central Processing Unit (CPU) alone, the memory allocation method reduces the execution time for memory copies by approximately 60% when a GPU is used, while utilizing read-only memories results in an approximately 20% reduction in the overall program execution time. Moreover, expanding the number of GPUs from one to four reduces the execution time by approximately 10%.