<div class="gmail_quote">On Wed, Apr 25, 2012 at 06:42, Thomas Witkowski <span dir="ltr"><<a href="mailto:thomas.witkowski@tu-dresden.de">thomas.witkowski@tu-dresden.de</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
Also some coarse space method. When not all ranks contribute to the
coarse space, a matrix defined on the coarse space leads to local
empty matrices on these ranks. So the question is if this leads to
technical troubles in PETSc. The share of nodes participating in the
coarse space may vary.</blockquote></div><div><br></div><div>There is no "technical trouble".</div><br><div>Communicator contraction doesn't really affect the MatMult. Depending on the hardware, it may affect the VecDot and VecNorm (contracting the communicator is likely worse on Blue Gene/P because certain reduction algorithms become unavailable---unless the contraction factor is massive and the remaining participants are nearby on the torus, it's probably good on many fat tree networks).</div>
<div><br></div><div>Redistributing the data for some locality (independent of whether you also contract the communicator) is more useful for MatMult, but you have to weight the benefits relative to the cost of moving the data, and it may not pay off.</div>