<div dir="ltr">Jed,<div>Let us discuss it tomorrow in the lab.</div><div>Hong</div></div><div class="gmail_extra"><br><br><div class="gmail_quote">On Tue, Jun 4, 2013 at 3:14 PM, Jed Brown <span dir="ltr"><<a href="mailto:jedbrown@mcs.anl.gov" target="_blank">jedbrown@mcs.anl.gov</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div class="im">Hong Zhang <<a href="mailto:hzhang@mcs.anl.gov">hzhang@mcs.anl.gov</a>> writes:<br>
<br>
> He needs MatTransposeMatMult_MPIDense_MPIAIJ<br>
> This requires huge amount of data communications among MPI processes.<br>
<br>
</div>In the original form, the result would have ten rows and 1400 Trillion<br>
columns, which isn't something reasonably storable in a MPIAIJ matrix.<br>
So the column space needs to be compressed, but then it still makes more<br>
sense to store the result matrix in transposed form (as ten columns)<br>
since those can be reasonably distributed. My MAIJ suggestion would<br>
work that way, as would MatTransposeMatMult_MPIAIJ_MPIDense.<br>
</blockquote></div><br></div>