AWS RADIUSS builds

[davidbeckingsale]

Adds stacks for a small subset of RADIUSS products on AWS

[vsoch]

🥞

[tldahlgren]

@davidbeckingsale Can you rebase to develop? This PR appears to be suffering from the bootstrap-related failures that have since been fixed.

[v-dobrev]

In this spec, it is better to also add ^hypre+cuda (no need to add cuda_arch after this since the mfem package propagates it to hypre).

[v-dobrev]

In this spec adding ^hypre+cuda causes issues, so we should not add that.

However, we can add ^superlu-dist+cuda cuda_arch=70 ^petsc+cuda.

[v-dobrev]

These two specs can be updated similar to the other updates I mentioned above.

[davidbeckingsale]

@v-dobrev seeing a couple of build failures - do you know if these are expected? I might revert to the builds without extra libs for this PR, and we can work on the others later.

[v-dobrev]

It looks like the PETSc + CUDA configuration step fails with errors like:

*******************************************************************************
         UNABLE to CONFIGURE with GIVEN OPTIONS    (see configure.log for details):
-------------------------------------------------------------------------------
--with-cuda-dir=/home/software/spack/[padded-to-384-chars]/linux-amzn2-aarch64/gcc-7.3.1/cuda-11.7.0-pe4amfjsidlxc2f4hbbvd7g32t2eyafo did not work
*******************************************************************************

on both aarch64 and graviton2.

Another build issue is another PETSc + CUDA configuration step that fails with

*******************************************************************************
         UNABLE to CONFIGURE with GIVEN OPTIONS    (see configure.log for details):
-------------------------------------------------------------------------------
Cannot use SuperLU_DIST with 64 bit BLAS/Lapack indices
*******************************************************************************

on x86_64_v3.

I did test the spec mfem +superlu-dist+petsc+sundials+cuda cuda_arch=70 ^superlu-dist+cuda cuda_arch=70 ^petsc+cuda on Lassen yesterday and it built fine.

Maybe we can just remove the ^petsc+cuda from the spec?

Another option is to just go back to the spec that worked before, removing the newly added ^superlu-dist+cuda cuda_arch=70 ^petsc+cuda.

[vsoch]

Moving forward this is going to be challenging because getting any spec right or to build is a moving target - I've found that previously working libraries break rather frequently either with some spack update or a dependency change. I don't know how to consolidate that aside from requiring regular builds.

[v-dobrev]

@vsoch, you make a good point. My understanding is that all Spack PRs require these checks to pass, so any new PR that breaks things should resolve any such issues. If that is not enforced then builds will remain broken until fixed. I'm not sure what the Spack policy in this regard is.

[davidbeckingsale]

I'll try the simpler spec and we can iterate in a future PR.

[vsoch]

@davidbeckingsale I'm not sure how to review this - I'd just say tests should be green?

[tldahlgren]

@v-dobrev All of the tests passed. Is this ready to be merged?

[v-dobrev]

My approval is primarily for the MFEM specs used. For the majority of the other additions, I'm not well qualified to review.

[tldahlgren]

My approval is primarily for the MFEM specs used. For the majority of the other additions, I'm not well qualified to review.

No worries. I suspect there will need to be tweaks along the way. Thanks.