Maintenance on December 20, 2021

We will be decommissioning /scratch2 and Jupyterhub starting 9:00 am Monday December 20th, 2021. We will also be making some changes to the servers hosting Open OnDemand, website and documentation at the same time. During this time Open OnDemand, website and documentation will be inaccessible. We expect the outage for Open OnDemand, website and documentation to take about 3 hours. After this outage Open OnDemand will no longer require Clemson VPN for access. We will send an update when Open OnDemand, website and documentation are back online.

Rest of Palmetto will not be affected.

Open OnDemand is a new web interface to Palmetto and provides functionality for deploying Jupyter servers in the same manner as JupyterHub.

Townhall meeting Aug 3rd 2021

After Palmetto maintenance Jul 2021, we hosted the Townhall meeting with all users to talked about changes to Palmetto after maintenance. The following points were discussed in detail:

  1. Introduction Open OnDemand
  2. 34 new A100 GPU nodes in phase 27
  3. New OMP_NUM_THREADS behavior
  4. Renaming of certain Modulefile names
  5. Autofs features
  6. Globus transfer file and scratch 1 reinitializing on Monday 08/09/2021 at 9am
  7. New Data transfer node. 

The recording and Q&A can be found in the following link (pls login using g.clemson.edu): bit.ly/CITI_Townhall_2021

Palmetto maintenance

Palmetto’s yearly maintenance is scheduled for the week of July 16th - July 23rd, 2021.

During the maintenance period there will be no access to Palmetto/JupyterHub/LoginVM.

Please retrieve any files you may need before the cluster goes offline.

All nodes and all storage will be down during the maintenance period.

Storage will be available when we bring the non-GPU nodes back online before benchmarking.

This year's plan for maintenance:

  • We are adding a new phase with new nodes.
  • Various Operating System Updates and Patches
  • Updating Spack and doing some housekeeping on the current packages installed.
  • Nodes in Phase 7a – 21 will be used for benchmarking. Phase 0 – 6 and storage will be brought back online before running the benchmark. We will send a notification once these systems are online and available.

Note:

1) /scratch1 will not be wiped but /scratch2 will be wiped. Please back up any data stored in your /scratch2 directories as it will not be recoverable after maintenance. 2) Any jobs that are started before maintenance with a walltime that conflicts with the start of the maintenance will not run.

If you have any questions or concerns, please reach out to us by emailing ithelp@clemson.edu with “Palmetto Maintenance” in the subject line.

April 30, 2021: a new version of XQuartz is released, so X11 tunneling now works for Mac OS > 10.14

Good news for Mac users! Recently, there has been an update to XQuartz, which we can use to set up X11 tunneling on Palmetto in order to run simple graphical applications. The older versions of XQuartz only supported Mac OS up to 10.14 (Mojave); the new version also supports Mac OS 10.15 (Catalina) and Mac OS 11 (Big Sur). We have added the guide to X11 tunneling to our website.

April 1st, 2021: Requesting "interconnect=any" and "gpu_model=any" in a qsub

When you request resources on Palmetto, you need to specify the type of interconnect (e.g. "fdr", "hdr", "1g"). If you have no preference about the type of interconnect, you can specify interconnect=any. Likewise, if you would like to request a compute node with a GPU but you have no preference for a particular GPU model (e.g. "k20", "k40", "p100", "v100"), you can specify gpu_model=any. This feature is also implemented in the JupyterHub interface.

For example:

qsub -I -l select=3:ncpus=4:mem=10gb:interconnect=any:ngpus=1:gpu_model=any,walltime=1:00:00

Here, we request 3 compute nodes with a GPU, but we don't care which GPU model and which interconnect type they have. We might end up with three nodes having different types of interconnect and different GPU models.

The benefit of this feature is that it increases the pool of nodes that fit your request. For example, if you need 20 nodes with a GPU, but don't care which particular GPU model they have, you won't need to have to wait for 20 nodes with the same GPU model to become available.

Please be aware of the differences in maximum walltime for different interconnect types (the walltime limit for interconnect=1g is 336 hours, whereas the walltime limit for interconnect=fdr and interconnect=hdr is 72 hours). If you request interconnect=any, you cannot request for more than 72 hours of walltime. If you need to request more than 72 hours, please use interconnect=1g.

Also, if you are using OpenMPI, it would be better to stick with interconnect=1g, interconnect=fdr, and interconnect=hdr, rather than use interconnect=any. The MPI modules are different on Ethernet nodes (1g) versus Infiniband nodes (fdr and hdr), so there is no guarantee that OpenMPI will run smoothly across nodes with heterogeneous interconnect type.

The old specifications will still work, for example

qsub -I -l select=3:ncpus=4:mem=10gb:interconnect=1g,walltime=200:00:00
qsub -I -l select=3:ncpus=4:mem=10gb:interconnect=hdr:ngpus=1:gpu_model=v100,walltime=1:00:00

There is no need to change your old scripts -- they will work as before.