<div class="gmail_quote">On Sat, Aug 27, 2011 at 17:09, Likun Tan <span dir="ltr"><<a href="mailto:likunt@andrew.cmu.edu">likunt@andrew.cmu.edu</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
<div id=":8s">Another question is, if the size is small in x and y directions, and large<br>
in z direction, i want more of processors, but the number is restricted to<br>
the size in x and y direction, since partition is not allowed in z<br>
direction, any way to improve efficiency?</div></blockquote></div><br><div>This is a fundamental problem. If there is tight coupling in the z direction, then it is algorithmically good to keep it on a single process, but it is not good for parallel scalability because you are not allowed to decompose in that direction. You can write a new integral function with the z direction decomposed, but it's significantly more complicated.</div>
<div><br></div><div>The practical solution is usually to not decompose in z and pay the price of somewhat reduced parallel scalability in exchange for simplicity, algorithmic performance, and efficiency.</div>