On 3/5/06, Florin Andrei <florin(a)andrei.myip.org> wrote: physical modeling. It still chews TONS of cpu. However, supercomputer means non-realtime music (realtime and mpi? i don't think so.) and you can do physical modeling offline fine on a normal machine. So well, not very much. Loki

On Sun, Mar 05, 2006 at 06:30:14PM +1100, Loki Davison wrote: There is more than MPI for super computing though. There are still vector machines, and there are still NUMA style machines like Origins, Altixs, or those large POWER or SPARC machines. I don't know that any of those are suitable for realtime either, but unless the answer is already known, they would be fun to explore for such purposes. Also, I wonder about clusters that don't use MPI. I seem to recall an organ system that used a cluster of PCs for physical modelling. Does anyone know how the different nodes communicated? I think it might have been something like master->node->audio output, where each node only modelled the pipes that it was responsible for outputing. -- Joshua D. Boyd jdboyd(a)jdboyd.net http://www.jdboyd.net/ http://www.joshuaboyd.org/

Loki Davison wrote: On NUMA, remote memory access is done in constant time (the so-called "NUMA factor"), so all memory accesses are bounded in time. Not deterministic, but the high bound is known. But even on cluster, MPI may be suitable for realtime audio. The typical latency of MPI over a high performance network is bellow 10 usec which should be enough for audio. -a.