Abstract
The Fast Fourier Transform (FFT) is one of the most widely used algorithms for scientific and engineering computation. Large FFTs are increasingly important in areas such as astronomy, fluid and molecular dynamics, image and signal processing and astrophysics. Our research seeks to improve portability of FFT codes while preserving high performance and scalability. The efficiency of many large-scale applications depends upon efficient implementation of FFTs, which may compromise 60-80% of the total computational work. With the emergence of Grids a scientist may use many different platforms for his/her research and portability of codes while preserving high performance become essential. Efficient FFT implementations are difficult to build because the low operation count per data element, one of the major reasons for the popularity of FFTs. The FFTs are very sensitive to memory systems and in case of parallel platforms the communication system since the FFT requires interaction between all variables. Furthermore portable implementations are difficult to build due to the growing heterogeneity of hardware architectures, complexity of memory systems and diversity of communication systems of parallel machines. Our approach to this challenge is self-adaptive codes, i.e. codes that adapt themselves to both the execution environment and the instance of the application at hand. We present the efficacy of an adaptive library based approach for building a scalable FFT framework. In line with the goal the major contributions of this research is an efficient and portable library for computing FFTs on parallel machines called MPI-UHFFT.

Date: Wednesday March 29, 2006
Time: 2:00 PM
Place: 218-PGH

Faculty, students, and the general public are invited.
Thesis Advisor: Prof. L. Johnsson