The new benchmarking effort aims to benchmark a variety of implementations of parallel FFT implementations for the purposes of the Exascale Computing Project (ECP). Many scientific applications of strategic importance to DOE need multi-dimension Discrete Fourier transforms, including ICL’s own heFFTe project. Over the course of the CLOVER project and more specifically its heFFTe component, the need for comparing FFT implementations became apparent to understand how the tradeoffs made by the developers affect the floating-point performance and mitigate communication overheads.
In its current incarnation, the FFT Benchmarking project released a technical report, ICL-UT-21-03, to provide an interim update on comprehensive testing and measuring various aspects of the FFT implementations for three-dimensional transforms on high performance systems. The results from the report reveal significant differences in functionality, performance, and supported hardware and give the ECP application community a much better appreciation of the available software choices.