| 摘要 |
<p>A plurality of computing processes including all-to-all (A2A) communication are optimally scheduled across a plurality of nodes (processors) constituting an n-dimensional (torus or mesh) network. The plurality of nodes (processors) constituting a network are partitioned into a communication (computing process) phase (A2A-L) requiring all-to-all communication for only among a plurality of nodes included in a first partial group, and a communication (computing) phase (A2A-P) requiring all-to-all communication for only among a plurality of nodes included in a second partial group, and the phases are made to overlap for parallel processing across a plurality of threads (thread 1, thread 2, thread 3, and thread 4). Parallel processing can be performed for such a plurality of computing processes as FFT (Fast Fourier Transform) and T (transpose) (internal transpose) as well.</p> |
