2022-04-22 Updates completed

Update started 2022-04-11, and now all nodes from phase 1 have been updated to Rocky Linux 8 (see explanation below). Jobs using the module tree needs to be updated for the new module tree and be resubmitted from alvis2 until the remaining nodes have been updated.

Virtual environments, or software built for CentOS 7 will likely not work. Containers should work without modifications.

Note that the foss/2021a and foss/2021b toolchains are not yet available due to a technical issue with CUDA support in the OpenMPI modules. Work in progress.

Alvis phase 2

The second phase of Alvis is finally entering production.

New hardware include many new GPU:s:

  • 8 nodes with 4 x A100fat (80GB VRAM)
  • 76 nodes with 4 x A100 (40GB VRAM)
  • 85 nodes with 4 x A40

You can find detailed hardware descriptions on the about Alvis page.

You can start requesting these nodes now. Some hardware issues are still being resolved and you can expect additional nodes to be released to the queue in the coming days.

New operating system and module tree

The new nodes are running Rocky Linux 8 and an updated Mellanox Infiniband stack. Together with the new architecture (Icelake) and operating system, a new module tree is required. You can view this from the new login node alvis2.

User installed software, like virtualenvs made for CentOS 7 may not work, and you should recreate them for phase 2 on the new login node.

New login node:

The new login node alvis2 has very fast internet and storage connectivity and is preferred for large data transfers. It does not have any GPUs.

Updating OS on phase 1

The hardware from phase 1 will be updated to Rocky Linux 8 in the coming weeks. They will get a new module tree mirroring the tree provided for phase 2. To simplify the procedure, the nodes will be drained and updated in blocks, redeploying all nodes and lastly the old login node alvis1.c3se.chalmers.se. Jobs waiting in the queue during the transition may fail (due to the OS change) and should be resubmitted from alvis2.

Mimer

The new storage system will be accessible under /mimer. The new A100 and A100fat nodes will have super fast access to this filesystem, removing the need for $TMPDIR. A40 nodes, and the V100 and T4 nodes from phase 1, will only have fast access.