Benchmark
Give an overview of the solver's performance.
Reference performance page is available here. Some results are also given in the last appendix of this PhD thesis.
Notes:
- GMRES uses a tolerance of 1e-5 and 50 restarts, and is preconditioned with an incomplete (fill-in factor of 2 and drop tolerance of 1e-6) LU matrix.
List of machines:
| Id | name | operating system | machine type | computing unit (frequency) | memory | compiler |
|---|---|---|---|---|---|---|
| 01 | - | ubuntu20 | virtual | i7-7700HQ (2.8Ghz) | 8Gb | GNU gcc |
| 02 | - | windows10 | laptop | i7-7700HQ (2.8Ghz) | 16Gb | VS 2022 |
| 03 | - | ubuntu20 | virtual | i7-4790K (4.0Ghz) | 8Gb | GNU gcc |
| 04 | - | windows10 | desktop | i7-4790K (4.0Ghz) | 12Gb | VS 2019 |
Test case:
The Onera M6 wing (dart/benchmark/onera.py) is the standard test/benchmark case for steady transonic flow computations.
Results
v1.2.0 (October 2022)
Mesh size: 711976 elements (556008 in the volume) and 135233 nodes.
| Machine id | Solver | iterations | threads | CPU_user | CPU_real | CPU_kernel |
|---|---|---|---|---|---|---|
| 01 | PARDISO | 8 | 1 (4) | 98 (122) s | 100 (45) s | 1.9 (3.0) s |
| 02 | GMRES | 8 | 1 (4) | 166 (223) s | 172 (96) s | 5.5 (8.1) s |
| 03 | PARDISO | 8 | 1 (4) | 105 (115) s | 78 (44) s | 5.0 (7.0) s |
| 04 | PARDISO | 8 | 1 (4) | 143 (199) s | 148 (84) s | 5.6 (7.1) s |
v1.0.1 (September 2021)
Mesh size: 772197 elements (602199 in the volume) and 146744 nodes.
| Machine id | BLAS | Solver | iterations | threads | CPU_user | CPU_real | CPU_kernel |
|---|---|---|---|---|---|---|---|
| 01 | Intel MKL | PARDISO | 15 | 1 (4) | 152 (202) s | 157 (79) s | 5.5 (8.5) s |
| 02 | Openblas | GMRES | 15 | 1 (4) | 191 (263) s | 195 (127) s | 3.5 (7.7) s |
| 03 | Intel MKL | PARDISO | 15 | 1 (4) | 138 (194) s | 142 (74) s | 5.5 (4.4) s |
| 04 | Intel MKL | PARDISO | 15 | 1 (4) | 204 (312) s | 211 (116) s | 6.4 (21.6) s |