On May 8, 2013, at 8:55 AM, Jed Brown <jedbrown at mcs.anl.gov> wrote:
> Barry Smith <bsmith at mcs.anl.gov> writes: > >> On May 8, 2013, at 7:54 AM, Peter Brune <prbrune at gmail.com> wrote: >> >> You guys are speculating too much. Create a DMDA characteristic >> for their code with reasonably large array dimensions and then >> -malloc_dump it (or something similar) to find out where the >> memory actually is. Then given this table of sizes propose >> changes, not before. > > I remember doing this before, maybe last time the issue was discussed > here. Flaky internet here is making it hard to find the thread. > Anyway, they want to have a total of two or three vectors that use all > available memory, thus they need to eliminate all inessential > problem-sized data. Fine, data is order M*N*P, but what about order M*N + N*P + M*P data? Some of the scatters are this size and hence ok to keep around? Barry
