<div class="gmail_quote">On Tue, Jun 7, 2011 at 21:12, Boyce Griffith <span dir="ltr"><<a href="mailto:griffith@cims.nyu.edu">griffith@cims.nyu.edu</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
In fact, it was the tests of assembled-versus-unassembled performance that you described on the libMesh list that led to me trying this out in the first place. My (possibly faulty) reasoning was that if there could be some performance advantage to using assembled matrices for low order FE discretizations, then perhaps there also might be some for low order FD/FV discretizations. At least in what I was doing, there did not appear to be any benefit to doing this, although matrix free is not tons faster either.</blockquote>
</div><div><br></div><div>For FEM with 3D hex elements, the assembled versus unassembled tradeoff is shown in the second figure here</div><br><div><a href="https://github.com/jedbrown/dohp/wiki/Dohp">https://github.com/jedbrown/dohp/wiki/Dohp</a></div>
<div><br></div><div>So assembled is good for scalar problems with lowest-order elements. As the order and number of components per node increases, unassembled becomes preferable. Note that this exploits a tensor product basis so the interfaces in libmesh, deal.ii, etc. will have much worse asymptotics as the order is increased.</div>
<div><br></div><div>For FD, you have an explicit formula for the local residual that does not involve quadrature and lots of element operations. This makes matrix-free residuals very cheap.</div><div><br></div><div>As perhaps an extreme example, for constant coefficients and a 27-point stencil, we got 93% of FPU peak in L1 and 71% from memory on Blue Gene/P. There are technical reasons why sparse mat-vec cannot saturate the memory bus on BG/P, but even if it could, there is 27*(1.5)/2 = 20 times more data to move through when you have a matrix.</div>
<div><br></div><div>I don't have a performance number for 27-point assembled, but 7-point assembled gets 7.0 Mstencil/s on BG/P. Our matrix-free stencil implementation gets 178 Mstencil/s in L1 and 71 Mstencil/s from memory. Note that it's much more likely for a given problem size to fit in L1 if there is no matrix involved, so the in-L1 results are possible. (Our implementation is not trivially obvious to the most casual observer, but it does demonstrate what the hardware can do.)</div>
<div><br></div><div>BG/P stencil computations:</div><div><a href="http://59a2.org/files/ppc450d_stencil_microkernel.pdf">http://59a2.org/files/ppc450d_stencil_microkernel.pdf</a></div><div><br></div><div>7-point MatMult numbers from:</div>
<div><a href="http://hpc.sagepub.com/content/early/2010/12/03/1094342010389857.abstract">http://hpc.sagepub.com/content/early/2010/12/03/1094342010389857.abstract</a></div><div>(<a href="http://www.mcs.anl.gov/uploads/cels/papers/P1658.pdf">http://www.mcs.anl.gov/uploads/cels/papers/P1658.pdf</a>)</div>