Docker ci build #507
Docker ci build #507
Conversation
TomMelt
force-pushed
the
docker-ci-build
branch
from
68bfd68
to
b79e0d0
Compare
TomMelt
force-pushed
the
xios-interface-with-tests
branch
from
bb5401d
to
b0a551d
Compare
TomMelt
force-pushed
the
docker-ci-build
branch
from
b79e0d0
to
235d3a3
Compare
* updated/tidied Dockerfiles * modified the CI to use the docker image hosted on ghcr.io * disabled XIOS and MPI for the mac build * added spack.yaml needed to generate dockerfile
TomMelt
force-pushed
the
docker-ci-build
branch
from
cd6731d
to
6cf186c
Compare
| . /opt/spack-environment/activate.sh | ||
| apt install -y wget | ||
| cd build | ||
| (cd core/test && wget "ftp://ftp.nersc.no/nextsim/netCDF/partition_metadata_1.nc") | ||
| for component in core physics | ||
| do | ||
| cd $component/test | ||
| # run serial tests | ||
| for file in $(find test* -maxdepth 0 -type f | grep -vP "_MPI\d*$"); do echo $file; ./$file; done | ||
| # run MPI tests that require 2 ranks | ||
| for file in $(find test* -maxdepth 0 -type f | grep -P "_MPI2$"); do echo $file; mpirun -n 2 --oversubscribe ./$file; done | ||
| for file in $(find test* -maxdepth 0 -type f | grep -vP "_MPI\d*$") | ||
| do | ||
| echo $file | ||
| ./$file | ||
| done | ||
| # run MPI tests |
There was a problem hiding this comment.
If the download in line 53 happens quickly (or could be put in a different step) then it might be nice to have separate steps for the serial tests and the MPI tests in the Dockerfile?
There was a problem hiding this comment.
I chose not to split because we can reuse more of the code that way. I don't have a strong preference though. I thought this was just easier to maintain. As for the download it should take about 1s.
There was a problem hiding this comment.
[Leaving thread unresolved because in today's meeting we decided to have separate workflows for MPI and non-MPI tests.]
| make | ||
| # added Python_ROOT_DIR to help cmake find correct Python in the mac VM | ||
| cmake -DENABLE_MPI=OFF -DENABLE_XIOS=OFF -DPython_ROOT_DIR=/Library/Frameworks/Python.framework/Versions/Current/bin .. | ||
| make -j 4 | ||
| - name: run tests |
There was a problem hiding this comment.
If we do go with separating serial and MPI-parallel tests for the Ubuntu tests as suggested above then it would make sense to rename this step as "run serial tests", or similar.
There was a problem hiding this comment.
as mentioned in the above reply I do not want to split at present because it will lead to some duplication.
There was a problem hiding this comment.
[Leaving thread unresolved because in today's meeting we decided to have separate workflows for MPI and non-MPI tests.]
Nextsim CI **now with added Docker**
This PR mainly pertains to the ubuntu build of the code.
Previously the CI workflow built
nextsimusing a GitHub VM instance that wouldapt installbasic dependencies required for the build process.However, with the addition of
xios(#496 ) this is no longer a tractable solution. Furthermore, it would take ~15 mins just to buildxiosbefore then going on to buildnextsim.Now I have created a
Dockerfileand associated pre-built image (currently on my Dockerhub:tommelt/nextsimdg-dev-env:with-xios) which contains all the prerequisites to build and test the code for our typical CI workflow.In theory it could also be repurposed by developers as a dev environment (although I personally don't love this).
more about the dockerfile
The
Dockerfileis based on aspackenvironment file (spack.yaml) was generated originally using the following command:However, I needed to make some manual adjustments to the
Dockerfile(necessary for installingxiosand some other dependencies).Dockerfiles/Dockerfileis the one that was used to generate the docker imagetommelt/nextsimdg-dev-env:with-xios.The image is
2.41GBand contains all prerequisites forMPIandXIOS. When uploaded to dockerhub (docker push tommelt/nextsimdg-dev-env:with-xios) it is compressed to625.97 MB.benefits of Docker in CI
nextsim🔥XIOSalone used to take ~15 minutesdocker pull tommelt/nextsimdg-dev-env:with-xios) and test stuff locallynextsimis necessarily adding complexity, which increases the required dependenciesChanges to the parallel (MPI) test naming convention
testRectGrid_MPI).testname_MPIXwhereX= the number of ranks the test should be run withdoctestspecifies minimum/maximum numbers of procs for the test to run. A test that requires 2 procs to run, run with 1 MPI proc instead will "pass" but with a message like "test skipped". This is not a situation we want to end up i.e., thinking our tests pass when they have actually not been run.To Do:
@timspainNERSC to fix MacOS build.I tried to fix MPI/xios build on mac (using tmate plugin to debug github action live) but the fix requires a lot of effort. (Compiling most/all dependencies from source) and it may not be worth it. We could consider not running the MPI/xios tests on mac? For now I have simply disabled MPI and XIOS on the mac build. One other work around would be create a dedicated spack build cache. (decided to not run tests for MPI/XIOS build on mac)docker build .)should we keep existing docker images? or "rebase" them? or just have one with all dependencies and users can build their own version of nextsim to meet their needs?dev environment build is sufficient for now.