It was just announced at the awards session at SC12 that Intel has won “best paper” for research into more efficient processing for a fundamental calculation in high performance computing, entitled “A Framework for Low-Communication 1-D FFT.”
Numerous wave applications (e.g. sound, radio) rely on the Fast Fourier Transform (FFT) including signal processing, communications, and multi-media. However, it is a very challenging problem to parallelize effectively. This is because for big FFT datasets running on large clusters, 50%-90% of time can be spent waiting on node-node data transfers rather than useful calculation.
Intel’s Software and Services Group & Intel Labs devised a new framework for distributed 1-D FFT problems which traditionally require three costly all-to-all inter-node data exchanges. The new approach delivers multiple 1D FFT algorithms requiring just a single all-to-all inter-node data exchange. According to the research, for large-scale problems this can double FFT performance (see the paper for details). Another key feature is that users can opt to further increase FFT performance by accepting reduced-accuracy results, so the algorithms scale to fit the needs of the particular application.