TauBench

This is an unstructured grid benchmark. The benchmark mimics the run-time performance of the Tau solver (a Navier Stokes solver which has been developed at the DLR in Germany). Tau is a three-dimensional parallel hybrid multigrid solver, which solves the Reynolds-averaged Navier-Stokes equations by finite volume scheme.

Install

$ make
mpicc -g -O2 flux.c -c
mpicc -g -O2 main.c -c
mpicc -g -O2 lim.c -c
mpicc -g -O2 smooth.c -c
mpicc flux.o main.o lim.o smooth.o   -lm -o taubench

To change the compiler and compiler flags

$ make 'MPICC = mpicc.openmpi' 'CFLAGS = -O3 -mtune=native -march=native -flto'
mpicc.openmpi -O3 -mtune=native -march=native -flto flux.c -c
mpicc.openmpi -O3 -mtune=native -march=native -flto main.c -c
mpicc.openmpi -O3 -mtune=native -march=native -flto lim.c -c
mpicc.openmpi -O3 -mtune=native -march=native -flto smooth.c -c
mpicc.openmpi flux.o main.o lim.o smooth.o   -lm -o taubench

Run

Tau supports both vector and cache colored grids, you might also need to adapt the compiler directives in nodep.inc and expand.inc. The benchmark itself can be used with two flags, the gridsize per process and the number of pseudo steps:

$ mpiexec -n 2 ./taubench -n 100000 -s 10
This is TauBench.
Evaluating kernels - please be patient.
..........

        - kernel_1_0 :      0.315 secs -   4470.269 mflops
        - kernel_1_1 :      0.131 secs -   1666.135 mflops
        - kernel_2_1 :      0.210 secs -   4733.000 mflops
        - kernel_2_2 :      0.186 secs -   6560.788 mflops
        - kernel_2_3 :      0.089 secs -   2759.613 mflops
        - kernel_2_4 :      0.069 secs -   7992.762 mflops
        - kernel_3_0 :      0.344 secs -   7952.712 mflops

               total :      1.311 secs -   9658.964 mflops

points     :     100000
steps      :         10
procs      :          2

comp       :      1.302 secs
comm       :      0.009 secs
comm ratio :      0.007

on one node

$ mpiexec --mca btl sm,self ./taubench -n 100000 -s 10
This is TauBench.
Evaluating kernels - please be patient.
..........

        - kernel_1_0 :      0.312 secs -   4509.797 mflops
        - kernel_1_1 :      0.139 secs -   1564.283 mflops
        - kernel_2_1 :      0.164 secs -   6071.223 mflops
        - kernel_2_2 :      0.156 secs -   7844.050 mflops
        - kernel_2_3 :      0.072 secs -   3411.729 mflops
        - kernel_2_4 :      0.080 secs -   6946.915 mflops
        - kernel_3_0 :      0.157 secs -  17392.511 mflops

               total :      1.106 secs -  22890.682 mflops

points     :     100000
steps      :         10
procs      :          4

comp       :      1.095 secs
comm       :      0.011 secs
comm ratio :      0.010