AMG is a parallel algebraic multigrid solver for linear systems arising from problems on unstructured grids.
Download the AMG2013 tar file (updated May 20, 2013)
How to Build
$ tar xf amg2013.tar
$ cd AMG2013/src
Edit Makefile for your environment
The executable can be found in the 'test' directory.
You can also 'make clean' or 'make veryclean' from the src directory.
How to Run
AMG requires specifying both the dimensionality of the problem and the scale, in MPI ranks.
In order to change the number of MPI ranks for a given problem size, it is also necessary to change the dimensionality of the problem. AMG uses the -P switch to set MPI parallelism, and the -n switch for changing dimensionality for each MPI rank.
For example, the small problem used for the Hopper baseline runs used 4x4x6=96 MPI ranks, with each rank solving a 150x150x150 problem. The total problem size is 96x150x150x150=324,000,000. To redefine the problem to use 8 MPI ranks, the total problem size needs to be adjusted to the cube root of (324,000,000/8), which equals 343. In this case, the run script would be changed to "-P 2 2 2" and "-n 343 343 343". It is reasonable to use a non-cubic solution to reach the same problem size, but it is desirable to keep it as cubic as possible.
There are three scripts corresponding to the small, large and extra large problems as defined for the Trinity/NERSC-8 run rules, run-<problem size>.sh. The small problem is sized to use ~100 GB of aggregate memory. The large problem is scaled to use 512 times more memory, and the extra large problem is scaled up by 10,000. The scripts are set up to use the Cray runtime, as such mppwidth specifies the number of MPI ranks for the baseline runs.
Results are written to stdout. Report "GMRES Solve wall clock time" and also report "Iterations."
The value of "Final Relative Residual Norm" must be less than 1.e-06
- Initial Trinity/NERSC-8 version 0.9 added to this web site
- README updated with instruction to also report Iterations