On Jun 22, 2010, at 4:21 PM, Jed Brown wrote: > On Tue, 22 Jun 2010 11:30:47 -0500, Barry Smith <bsmith at mcs.anl.gov> wrote: >> Actually we should add the same support for matrices and then the >> inconsistency would disappear. >
Because we got an email from an intelligent person who went down the wrong track because of this inconsistency I have implemented this. Now the MatView() also remaps back to the natural ordering. Barry > Is this actually useful? I.e. would anyone actually want to work with > the matrix in natural ordering? If you read it in parallel, you'll get > "slabs" instead of blocks". If you have a different number of processes > than you started with, the old blocks at least provide a quasi-sensible > partition. Even if you read it in serial, the standard partition should > still have better memory locality. > > But perhaps the issue is not performance, but just being able to easily > compare matrices assembled on a different number of processes. That > does seem useful, but it seems like a fairly involved thing to > implement. I think I'd rather just put 1:N in the vector (in the > natural ordering) and apply that permutation from Matlab. > > Jed
