On 18 June 2015 at 18:48, Stefan Karpinski <[email protected]> wrote:
> On Wed, Jun 17, 2015 at 4:38 PM, Daniel Carrera <[email protected]> > wrote: > >> An irritating problem with all the codes that solve planetary systems is >> that they are all serial -- this problem is apparently hard to parallelize. > > > That appears to be a language-independent issue. Even with mature > threading, it doesn't seem like this problem would scale well. It might be > possible to figure out a clever way to parallelize it, but that is a > serious algorithmic research problem. A superficial Google Scholar search > indicates that some work has been done in this direction using tree data > structures. > Yeah, this is an algorithmic issue and many people much smarter than me are working on it. I was just hoping that using shared memory would give me a small amount of parallelism without clever algorithms. Tree codes are the standard solution for systems like galaxies, where the number of particles is very large, but the dynamical age is small. In a galaxy you have 200 billion particles that have each done 40 orbits. For planet systems the problem is reversed. We have few particles and long ages (Earth has gone around the Sun 4 billion times). This means that small errors can really accumulate. For these problems the state of the art is either very high order integrators, or "symplectic" integrators which are designed to preserve a Hamiltonian that is closely related to the Hamiltonian of the system. Unfortunately, unlike the tree codes, none of these solutions are parallel. Because every particle affects every particle and you can't afford to throw away any information, the problem is very resistant to parallelization. I was wondering if I could get a free lunch by using shared memory. Cheers, Daniel. -- When an engineer says that something can't be done, it's a code phrase that means it's not fun to do.
