Home

Welcome to the TeaLeaf_ref wiki!

4/11/14

Adding new test problems with pass criteria.

Test 2 takes about 20 seconds on a 16 core Sandy bridge, running 4 omp 4 mpi.

I increased eps to 1.0e-15 and it takes 33 ish.

Flat omp is slow so I need to look at first touch and numa issues.

13/11/15

Commit 338 on branch dev/deflation is functionally correct provided we uncomment: "inner_use_ppcg = .TRUE.". The code performance is limited by the costs of the internal halo exchanges between patches, as we increase the size of the deflation space i.e. tiles_per_task*tasks. There is a sweet spot where the improved performance of the compute on the fine grid and the reduced number of outer iterations is counterbalanced by the halo exchange and coarse grid solves.