An elastic-wave stencil is decomposed into row slabs of a grid

PHYSICS · Multi-part · Fermi University Physics · Computational Physics · High-performance and data-driven computational physics · High-performance and data-driven computational physics

Scientific Computing for Physicists, Ch Many processors: speed-up, decomposition and the cost of communication

Problem

An elastic-wave stencil is decomposed into row slabs of a Ng=2048N_{\rm g}=2048 grid. A ping-pong test gives αlat=0.8 μs\alpha_{\rm lat}=0.8\,\mu\mathrm{s} and βnet=10 GB s−1\beta_{\rm net}=10\,\mathrm{GB\,s^{-1}}, and one halo row carries NgN_{\rm g} double-precision values. Each process exchanges two such rows per step, and the time to update one cell on a fully loaded node is tcell=30 nst_{\rm cell}=30\,\mathrm{ns}. ((a)) Find the slab size n∗n^* (in cells owned by a process) at which the two halo messages cost as much as updating n∗n^* cells. ((b)) Find the value tcellt_{\rm cell} would have to take for that crossover to rise to exactly one row of the grid (Ng=2048N_{\rm g}=2048 cells). ((c)) State the factor by which tcellt_{\rm cell} would have to change from its given value of 30 ns30\,\mathrm{ns} to reach that crossover.

Browse the problem library