Hello,
first I agree a variable smoothing length is the better choise in a SPH algorithm. As I understood, one of the bigger problems is the implementation of the particles in Carpet, because of MPI. For using a tree structure to save and search particles there is a global memory outside of the parallelized threads necessary. In this case we could use variable smoothing length without problems, I think (it is still possible to make a search routine without trees, using the grid structure, but therefor it is also necessary to determine the grid cell enclosing a particle). -Stefan Am 2015-11-25 21:53, schrieb Erik Schnetter: > Stefan > > There exist several algorithms for managing the SPH particles and finding > interactions. Personally, I dislike using an algorithm that would e.g. > changing the particle radius depending on the process decomposition. > > If I was to implement an SPH method for astrophysics in Cactus, then I would > probably choose a binary tree. I would make the process decomposition that is > implied by this tree independent of how Cactus decomposes its grid functions. > Thus you can test both grid functions and particle algorithms independently. > > To couple particles and grids, you need two ingredients: > - Interpolate grid quantities at particle locations > - Deposit particle quantities onto grid > > The first already exists in Cactus; you would use the Cactus interpolator for > this. > > The second is particle-specific, and this routine needs to be written. > Determining the grid cell enclosing a particle is the main ingredient. For > PUGH (a uniform grid) this is straightforward; for Carpet, there is a routine > "gh::locate_position" that one would call. > > Apart from these considerations, I have a personal preference for algorithms > that are derived from a Lagrangian. A variable smoothing length is likely > important in astrophysics since you will encounter large density differences > there. > > -erik > > On Wed, Nov 25, 2015 at 2:18 AM, Stefan Ruehe > <[email protected]> wrote: > >> Good morning, >> >> I have made some thoughts about the problems with SPH and MPI. >> >> I found a paper by Valdez-Balderas et al (2012) >> (http://adsabs.harvard.edu/abs/2012arXiv1210.1017V [1]), which could help to >> solve some of the problems. >> >> They suggest a particle halo on each processor unit, in which the neighbour >> particle of the adjacent processors are saved. This should be synchronized >> in each timestep. >> >> One problem of this method is SPH have to use fixed smoothlength, otherwise >> the volume of the halo can't be set. The variability of the smoothlength is >> required to have an adaptive refinement in the SPH-algorithm. I would >> suggest to use semi-fixed smoothlength, which are smaller in higher >> refindement levels. This could reduce the disadvantages of the fixed >> smoothlength. >> >> What is your opinion this? >> >> Now I try to test how good SPH-approximations for the hydrodynamic grid >> variables in the Tmunu base are under the condition of such "adaptive-fixed" >> smoothlength. I have an other method in mind, but this would need more >> temporary memory. >> >> Best regards, >> >> Stefan Ruehe >> >> _______________________________________________ >> Users mailing list >> [email protected] >> http://lists.einsteintoolkit.org/mailman/listinfo/users [2] > > -- > > Erik Schnetter <[email protected]> > http://www.perimeterinstitute.ca/personal/eschnetter/ [3] Links: ------ [1] http://adsabs.harvard.edu/abs/2012arXiv1210.1017V [2] http://lists.einsteintoolkit.org/mailman/listinfo/users [3] http://www.perimeterinstitute.ca/personal/eschnetter/
_______________________________________________ Users mailing list [email protected] http://lists.einsteintoolkit.org/mailman/listinfo/users
