Samples installation in CUDAcore causes failure

[branfosj]

During the install of CUDAcore 11.3.1 the installation of the samples has the following error:

[ERROR]: Failed to update permissions for /rds/bear-apps/devel/eb-sjb-up/tmp/CUDAcore/11.3.1/system-system//NVIDIA_CUDA-11.3_Samples/1_Utilities

(that is the buildpath).

This causes CUDA to remove everything installed. I've tested with a different buildpath (GPFS storage, /dev/shm, and local disk /scratch) and that also fails.

If I do not install the samples (and removing the sanity check) then the installation completes.

[Micket]

Is this unique to POWER?

[branfosj]

No. I'm seeing it on both x86_64 and POWER. I already had CUDAcore-11.3.1.eb installed on x86_64 from before the installation of the samples was added to the CUDA easyblock (#2374).

[branfosj]

I'm seeing the same issue with:

• CUDAcore-11.3.0.eb
• CUDAcore-11.2.2.eb
• CUDAcore-11.2.1.eb
• CUDAcore-11.1.1.eb
• CUDAcore-11.0.2.eb

[Micket]

eb CUDAcore-11.3.0.eb --include-easyblocks-from-pr 2374 --rebuild

using latest eb from develop I end up with a working installation (with samples). Build dir is on a ramdisk.

[branfosj]

I've solved my issue. If I install inside a slurm job the cgroup causes problems with the installation of the samples. If I direct log on to the node then these install fine.