WIP: Ci docker use external mpich #3876
Conversation
- change base image (smaller image and newer tools)
- clone spack branch and configure mirror to match
- update packages file:
- so spack finds the external mpich
- so mgard matches the required version (now an adios2 dependency)
- fix external mpich install so the mpicc compiler actually works
- need the spack hwloc added to the mpich rpath
- may need same with other deps it should share with adios2 ci envs
- remove adios2-ci-deps environment and extra mgard install
- adios2 already has it
- but spack still concretizes two hashes of mgard
- the two mgards project to one view location and spack dies
|
It's odd that using the |
|
|
||
| COPY packages.yaml /etc/spack/packages.yaml | ||
|
|
||
| # Install Base specs | ||
| RUN . /spack/share/spack/setup-env.sh && \ | ||
| RUN git clone --depth 1 --single-branch --branch e4s-${E4S_VERSION} https://github.com/spack/spack && \ |
There was a problem hiding this comment.
spack should be already installed from the ecpe4s image, is there any reason for this clone?
There was a problem hiding this comment.
Spack is not already installed with the new image I used as a base.
There was a problem hiding this comment.
The new base image is used in spack ci as one of the runner images, and is intended to have spack cloned before each run.
There was a problem hiding this comment.
I see, any reason to use the runners image rather than the minimal spack image as we used before?
There was a problem hiding this comment.
Several reasons:
- the spack in the image we used before failed to concretize the stack with the external mpich
- according to the maintainer of the E4S images, the runner images are:
- better: modern gnu make, patchelf, cmake, ccache, ninja exposed in default PATH
- smaller:
1.91GBuncompressed
| # # configure with "--enable-g=dbg,log" for runtime debug output | ||
| # # | ||
| RUN . /spack/share/spack/setup-env.sh && \ | ||
| export HWLOC_LOCATION="$(spack location --install /$(spack find --format {hash:7} hwloc))" && \ |
There was a problem hiding this comment.
Why not spack load hwloc before this cmd?
There was a problem hiding this comment.
I'll try that and see if it allows me to remove the specification of the hwloc path in the configure args.
|
I made a mistake when building the images, forgetting to switch the I'm rebuilding/pushing now, and we'll see if it's any better. |
|
@scottwittenburg I guess that this was superseded by #3883, shall we close? |
|
Yes, thanks for the reminder. |
Build a non-spack mpich using
ch3:sock:tcp, which spack doesn't support, and work it to spack environment in the ci image as an external. Thech3:sock:tcpbuild should handle the oversubscription in CI much better, resulting in faster tests when building with mpich.