From: Martin Kronbichler Date: Fri, 17 Feb 2017 20:01:11 +0000 (+0100) Subject: Add picture of performance X-Git-Tag: v8.5.0-rc1~106^2 X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=c2103365f544b713ee80c44ed3a6d7c9cf7b6782;p=dealii.git Add picture of performance --- diff --git a/doc/doxygen/images/fe_evaluation_laplacian_time_per_dof.png b/doc/doxygen/images/fe_evaluation_laplacian_time_per_dof.png new file mode 100644 index 0000000000..e7d418017b Binary files /dev/null and b/doc/doxygen/images/fe_evaluation_laplacian_time_per_dof.png differ diff --git a/include/deal.II/matrix_free/fe_evaluation.h b/include/deal.II/matrix_free/fe_evaluation.h index c904aedadf..05614fc54a 100644 --- a/include/deal.II/matrix_free/fe_evaluation.h +++ b/include/deal.II/matrix_free/fe_evaluation.h @@ -1728,6 +1728,23 @@ protected: * case with compile-time bounds given through fe_degree selects another * template class without compromising efficiency. * + * An overview of the performance of FEEvaluation is given in the following + * figure. It considers the time spent per degree of freedom for evaluating + * the Laplacian with continuous finite elements using a code similar to the + * step-37 tutorial program. The time is based on an experiment on a single + * core of an Intel Xeon E5-2687W v4, running at 3.4 GHz and measured at + * problem sizes around 10 million. The plot lists the computational time + * (around 0.1 seconds) divided by the number of degrees freedom. + * + * @image html fe_evaluation_laplacian_time_per_dof.png + * + * The figure shows that the templated version is between 2.5 and 3 times + * faster. The fastest turnaround on this setup is for polynomial degree 5 at + * 7.4e-9 seconds per degree of freedom or 134 million degrees of freedom per + * second - on a single core. The non-templated version is also fastest at + * polynomial degree 5 with 2.1e-9 seconds per degree of freedom or 48 million + * degrees of freedom per second. + * *

Handling multi-component systems

* * FEEvaluation also allows for treating vector-valued problems through a