You may continue to submit jobs until the maintenance period begins, but if the system determines your job will not have time to finish, it will not start until Rivanna is returned to service. While drive mapping and project storage will be unavailable, other storage will remain accessible through Globus.

All systems are expected to return to service by Thursday, May 30 at 6 a.m.

IMPORTANT MAINTENANCE NOTES

Hardware and partition changes

1. afton: We are pleased to announce the addition of 300 nodes, 96 cores each, based on the AMD EPYC 9454 architecture. For the time being, these new nodes are placed into their own partition afton, in which each user can request up to 6000 cores in aggregate. There is some chance that 100 nodes will be available initially, in which case the core limit will be adjusted to 2000. (The full release is scheduled for July, where the nodes will be integrated into the other existing partitions. More information to follow.)
2. interactive: The dev and instructional partitions will be combined into interactive, which will have a maximum time limit of 12 hours and a maximum core limit of 24. The SU charge rate will be 1. Two RTX 2080Ti nodes (10 GPU devices each) will be added from the gpu partition.
3. gpu: Four NVIDIA A40 (48 GB) nodes, 8 GPUs each, will be added to the gpu partition. The RTX 2080Ti nodes (2 x 10 GPU) will be moved into the interactive partition.

Please contact us if you have questions about the new partitions or if you experience issues with the new hardware.

System upgrades

1. Operating System: Rocky 8.7 → 8.9. There is no need to rebuild your own code. (Intel users may need to do so but for a different reason; see below.)
2. Slurm: 21.08.8 → 23.02.7. Job-related commands remain the same.
3. NVIDIA driver: 535.104.12 → 550.54.14.

Modules

• Attn Intel users: With the addition of the Afton nodes based on the AMD EPYC architecture, we have reorganized and rebuilt all modules under the intel toolchain. If you used -x (e.g. -xavx) to build your own code, you should rebuild it with -march=skylake-avx512 for it to run on both AMD and Intel hardware. Below we list all the modules that are upgraded or moved to a different toolchain. (Intel modules not listed can be loaded the same way as before.) The toolchain needs to be loaded before the module, e.g. module load gcc gmp. The gompi toolchain is equivalent to gcc openmpi. There is no impact on the existing modules built with GCC.

• Attn NVHPC users: The compiler toolchain nvhpc and nvompi will be upgraded to 24.1 and 24.1_4.1.6, respectively. The previous versions 23.7 and 23.7_4.1.4 will be removed. All modules under this toolchain will be rebuilt. There should be no need to rebuild your own code.

• The following modules will be removed from Rivanna during the maintenance period.