세미나 및 강연

Data movements, when not carefully optimized, can easily dominate the execution time and energy consumption of running data-intensive applications. This is particularly true with modern computing systems with increasingly high degree of concurrency. Therefore, software that minimizes the data movements will play a key role in achieving desired parallel scalability and energy efficiency. In this talk, I will share my experience on how system software such as compiler and kernel library can economize the data movements significantly.
First, I will introduce Elk streaming programming system that optimizes the capacity of buffers in pipeline-parallel applications. The Elk system can achieve an order of magnitude smaller memory footprint than the previous state-of-the-art, significantly reducing on-chip and off-chip data movements. Second, I will describe segment-of-interest FFT algorithm that reduces three all-to-all data transfers required in large-scale cluster-level 1-D FFTs to just one, achieving up to 2x speedup compared to Intel MKL.