ABI DIFF from rttov_test

[Andrew Prata]

Hi James

I have just attempted a fresh install of RTTOV and run into something I haven’t seen before. When I run ./test_rttov14.sh ARCH=gfortran-conda, I see a DIFF for ABI. See below. Not a current issue for me as I’m not using ABI but just wondered if you’ve seen this before?

abi/501 96 6 1 70 XXX..2 co2 o3 20.31X..4 XX.XXXX.XX.321 1 1 0 1 1 …..XXXXX…. 0.46 0.24 DIFF

| ========== PROFILES_K( 94)%T ==========
| 8
| test_rttov14.1.gfortran-conda/abi/501/out/k/profiles_k.txt: 0.122215E-006
| test_rttov14.2/abi/501/out/k/profiles_k.txt : 0.122212E-006
+—————————————————————————————–

[James Hocking]

Hi Andrew,

It is not uncommon to see small inter-compiler/inter-platform differences like this. The vast majority occur in the adjoint/Jacobian model outputs which are very sensitive to tiny differences in the way calculations are performed by the compiler/CPU. In this case the output is reporting that the temperature Jacobian for one profile for one channel in one layer differs by 3.E-12 K/K (the difference between the two numbers) between your run and our reference test run. This difference is not scientifically significant. In this instance, the absolute value of this Jacobian element is ~1.E-7 K/K which is already extremely small (a 1 K change in the temperature in this layer would give a 0.1 micro-Kelvin change in BT).

Such differences are not specific to ABI, and depend on many factors including the specific input profile, configuration of the simulation, compilation flags, compiler version, and architecture on which the code is running.

Best wishes,
James

[Author]

Posts