Eric, we would welcome an implementation that batched together a run-time-specified number of RHS vectors and used MUMPS' native multiple-RHS routines. The problem is that there has to be enough memory to store them all on rank 0, which is non-scalable, but the MUMPS developers do not care to fix that aspect of their interface. The current implementation solves only one at a time, which is less efficient than solving several at a time.<div class="gmail_extra">
<br><br><div class="gmail_quote">On Thu, Dec 6, 2012 at 10:20 AM, Hong Zhang <span dir="ltr"><<a href="mailto:hzhang@mcs.anl.gov" target="_blank">hzhang@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">
Eric :<br>
<div class="im"><br>
> I just want to know if there is a known difficulty or unfeasible task if<br>
> someone would like to implement MatMatSolve_MUMPS?<br>
<br>
</div>MatMatSolve_MUMPS is supported in petsc-dev.<br>
It calls MatMatSolve_Basic() with then calls MatSolve_MUMPS<br>
in a loop.<br>
<br>
MUMPS own matmatsolve requires entire rhs matrix store in the host<br>
which is non-scalable.<br>
<span class="HOEnZb"><font color="#888888"><br>
Hong<br>
</font></span></blockquote></div><br></div>