SDSC: PKG - expanse/0.17.3/gpu/b - Missing Amber GPU(example application) #50
Comments
|
I still need to retry a build for amber22 on GPU. |
mkandes
changed the title
|
@mkandes - Jerry has an updated Amber Spack package that includes Python extensions with miniconda environment. |
|
Build still failing ... |
mkandes
added a commit
that referenced
this issue
Sep 20, 2023
We've encounted persistent build issues when utilizing CUDA 11.2.2 to build Amber 22 with GPU support using Spack on Expanse [1]. To get around these issues, we're deploying this compiler and package dependency chain into the expanse/0.17.3/gpu/b instance to support builds with CUDA 10.2.89. We expect reverting to the older compiler and CUDA versions may resolve the issue with Amber 22 as the deployment on TSCC2 required such a fix as well [2]. [1] #50 [2] [mkandes@login1 ~]$ module spider amber/22-4zxr2dr ---------------------------------------------------------------------------- amber: amber/22-4zxr2dr ---------------------------------------------------------------------------- You will need to load all module(s) on any one of the lines below before the "amber/22-4zxr2dr" module is available to load. shared gpu/0.17.3 gcc/8.5.0-mf5bqu2 intel-mpi/2019.10.317-f7l5rk4 Help: Amber is a suite of biomolecular simulation programs. Note: A manual download is required for Amber. Spack will search your current directory for the download file. Alternatively, add this file to a mirror so that Spack can find it. For instructions on how to set up a mirror, see http://spack.readthedocs.io/en/latest/mirrors.html [mkandes@login1 ~]$
|
Amber22 with GPU-accelerated support has consistently failed to build successfully on Expanse with CUDA 11.2.2. As such, we've decided to deploy an older gcc 8.x compiler and package dependency chain to support CUDA 10.2.89 within the expanse/0.17.3.gpu/b instance. |
|
Build succeeded. |
|
|
@nwolter - Ready for testing. |
|
@mkandes I assume you ran this as well? I am getting an error. Currently Loaded Modules:
Where: /cm/local/apps/slurm/var/spool/job25332963/slurm_script: line 21: pmemd.cuda: command not found |
|
Need to load amber/22 specifically. Example script can be found at: |
|
Build spec and standard output committed back to deployment branch. |
|
amber@22 for expanse/0.17.3/cpu/b would no longer build with the updated Spack package used to deploy amber@22 in expanse/0.17.3/gpu/b. |
|
The issue may have been that the python version installed by the latest miniconda installer was no longer compatible with the AmberTools23. Note, however, this is not entirely clear if this is the problem. |
|
Attempting to restart from the spack/spack upstream was not an option either as they have still not implemented the cmake-based build process for amber@22. As such, we now have at least 3 or more separate amber packages required to support all of the versions currently deployed on SDSC systems. We need to reconcile them into a general Spack package and submit these changes upstream to the Spack project. |
mkandes
added a commit
that referenced
this issue
Oct 3, 2023
After successfully using one of the latest amber Spack packages developed by Jerry Greenberg here at SDSC to deploy a GPU-accelerated version of Amber22 and AmberTool23 into expanse/0.17.3/gpu/b, it was found this package was incapable of redeploying the same CPU-only version into expanse/0.17.3/cpu/b. As such, we've had to revert to the original configuration of the package used to deploy into expanse/0.17.3/cpu/b. We'll eventaully need to reconcile these issues as neither of these packages support older versions of Amber either. See [1] for more information. [1] #50
|
Successfully redeployed and tested amber@22 for expanse/0.17.3/cpu/b. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
No description provided.
The text was updated successfully, but these errors were encountered: