Verify equivalent resolutions for code comparison
Created originally on Bitbucket by kburns (Keaton Burns)
Hello,
I'm Keaton Burns, one of the developers of Dedalus. We've seen the performance comparisons in the FluidSim paper and are interested in figuring out how we can improve our performance on fully-Fourier problems, but first wanted to verify some of the parameters of the tests being run. In particular, this is a follow-up to Pierre's posts over on our issue tracker here.
Our question is if the stated resolution for each of the test problems in your comparisons (e.g. 512^2) refers the number of grid points or the number of dealiased modes being used in the simulation.
It seems like the other codes in the comparison may have you specify the grid resolution and then use a 2/3 truncation for dealiasing. For a nominally 512^2 simulation, this would result in ~341^2 actual degrees of freedom (fully-dealiased modes). Is this the right interpretation of the specified resolution for FluidSim?
In Dedalus, the specified resolution is the number of dealiased modes and a 3/2 padding is applied for dealiasing. A Dedalus simulation nominally of 512^2 resolution, which it seems was used for the comparisons here, will therefore use a grid size of 768^2 for the nonlinear products, etc. We're wondering if such a discrepancy in the codes exists and may account for some of the observed performance gap.
Thanks! -Keaton