The CG solver in IPPL based on finite differences does not communicate halo cells properly. The relative error stays constant across all grid sizes.
I created a new test file for each test run here, I'll add them in the PR. Each test was run with 4 CPUs. Single Rank the results are as expected and converge to the correct solution.
TestCGSolver_convergence_constant:
size,h,relError,trueResidual,solverResidual,itCount,solveTime
4,0.2,0.5140887128517535,3.820123950655526e-16,4.878124942680895e-16,6,0.0009340000000000008
8,0.1111111111111111,0.6288001761049367,3.450766361756485e-15,6.512311282577119e-13,25,0.0002620000000000001
16,0.05882352941176471,0.6976987791335945,9.667134093282123e-14,5.012346698362658e-11,50,0.0008530000000000013
32,0.0303030303030303,0.735402974261647,8.992190749713994e-14,1.160814327987606e-10,101,0.006244999999999999
64,0.01538461538461539,0.7550241557554113,2.101455100276474e-13,3.71812280368397e-10,205,0.07481400000000001
128,0.007751937984496124,0.7650073441869345,1.029147810596784e-12,1.046144106420834e-09,414,1.095381
TestCGSolver_convergence_periodic:
size,h,relError,trueResidual,solverResidual,itCount,solveTime
4,0.25,0.6377366121760882,1.596057071843241e-16,5.123796534383003e-14,1,0.0008059999999999994
8,0.125,0.06610999025353599,4.70268102220917e-16,5.55931194985223e-14,4,8.000000000000021e-05
16,0.0625,0.01511692691447543,5.860301855464142e-11,1.112753509887523e-07,12,0.0003029999999999995
32,0.03125,0.003724276265085818,5.170018915267303e-11,2.776629330922694e-07,21,0.001708999999999999
64,0.015625,0.0009277485090611948,9.456697119791474e-11,1.436507970115203e-06,41,0.016218
128,0.0078125,0.0002317313213305733,8.187506446496449e-11,3.51770745253475e-06,81,0.225149
The first step is worse, as the resolution of the grid is to coarse to properly show the SinSin waves.
The CG solver in IPPL based on finite differences does not communicate halo cells properly. The relative error stays constant across all grid sizes.
I created a new test file for each test run here, I'll add them in the PR. Each test was run with 4 CPUs. Single Rank the results are as expected and converge to the correct solution.
TestCGSolver_convergence_constant:
TestCGSolver_convergence_periodic:
The first step is worse, as the resolution of the grid is to coarse to properly show the SinSin waves.