Exact numbers of iterations are sensitive to details of computations, so you need to check<div><br></div><div> a) that the solvers are exactly the same using -ksp_view</div><div><br></div><div> b) the convergence history, which can be different in parallel due to non-commutativity of floating point arithmetic</div>
<div><br></div><div> Matt<br><br><div class="gmail_quote">On Thu, Sep 23, 2010 at 7:51 AM, Leo van Kampenhout <span dir="ltr"><<a href="mailto:lvankampenhout@gmail.com">lvankampenhout@gmail.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">Hi all,<br><br>With p number of processors in the communicator, the block preconditioner PC_BJACOBI will by default use p blocks. So far, so good. However, in order to compare this algorithmic efficiency decrease (since the bigger p, the less efficient the preconditioner), i ran the commands<br>
<br><span style="font-family:courier new,monospace">mpirun -n 1 ./program -pc_bjacobi_blocks 8 </span><br><span style="font-family:courier new,monospace">mpirun -n 8 ./program -pc_bjacobi_blocks 8 </span><br><br>I expected the preconditioning to be equally efficient in this case. However, GMRES makes more iterations in the first case (30 against 28) which I cannot explain. Are there more subtle differences about the preconditioner or the KSP that i'm overlooking here?<br>
<br>regards,<br><font color="#888888"><br>Leo<br><br><br><br><br><br>
</font></blockquote></div><br><br clear="all"><br>-- <br>What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.<br>-- Norbert Wiener<br>
</div>