Modify verification tests

Open rfvander opened this issue 8 years ago • 1 comments

I just came across an issue with the stencil code (Fenix version, using OpenMPI 1.7.1 on a 4-node, 64 rank Xeon cluster). Problem size was 3584 and the number of iterations 2560. No errors injected. The test failed twice in a row: ERROR: L1 norm = 5121.999940, Reference L1 norm = 5122.000000 ERROR: L1 norm = 5121.999999, Reference L1 norm = 5122.000000 In both cases the relative error was less than 1.e-8, but the absolute error exceeded that limit. While it is conceivable there is an error in the runtime (the same run on a 36-core HSW system succeeded), the error is small enough that it should have passed. I feel we should update all the verification tests to measure relative error, not absolute.

Aug 20 '17 05:08 rfvander

Assignment accepted

Oct 27 '17 16:10 rfvander