Math HPC
Resources available
| Hostname (add .math.umd.edu) | OS | GPU | CPU | RAM |
|---|---|---|---|---|
| mathhpavw1202a | Red Hat 8 | 1 × NVIDIA Ampere A100 (80 GB) | Intel Xeon 6258R @ 2.7 GHz (112 threads) | 754 GB |
| mathhpavw1202b | Red Hat 8 | 8 × NVIDIA RTX A6000 (48 GB) | AMD EPYC 7763 @ 2.45 GHz (256 threads) | 4 TB |
| mathhpavw1202c | Red Hat 8 | 8 × NVIDIA RTX A6000 (48 GB) | AMD EPYC 7763 @ 2.45 GHz (256 threads) | 2 TB |
Requesting access
- Email math-helpdesk@umd.edu to request access.
- If you aren't listed as having a Math affiliation in the campus directory, include who you are working with in Math.
Connecting to the servers
If you aren't familiar with Linux, start with this introduction.
- If you are off campus, connect to the UMD GlobalProtect VPN first.
- Use Secure shell access to connect to your preferred server from the table above.
- Use SFTP to transfer files to and from the servers:
Errors about running out of processes
If you see errors related to running out of processes, you may have been enrolled in a CS or ECE class that instructed you to add a process limitation to your profile. This will cause problems with VS Code and other remote editors.
Go to your home space and check .cshrc.mine (most likely) and .bashrc.mine (less likely, but worth checking):
cd ~ vi .cshrc.mine
Look for a line containing ulimit, limit maxproc, limit process, or something similar. Comment it out by adding a # at the beginning of that line, then save and exit. Log back in and the issue should be resolved.
Storage
On mathhpavw1202a, b, and c there is a significant amount of local storage available to each user under /export/<username>. This storage is local to each server, meaning you can't see the storage on one server from any of the others. This local storage is not backed up and should be used as temporary space during your computations.
These servers also connect to a file server that provides a small amount of home space accessible from any TerpConnect server, so it can be seen from all three of these servers as well as glue.umd.edu and login.math.umd.edu.
You can see how much space is left with:
df -h
You should see something like the output below. Look for the / entry — the last one here — which has 1.7 TB free.
mathhpavw1202a:~$ df -h Filesystem Size Used Avail Use% Mounted on devtmpfs 377G 0 377G 0% /dev tmpfs 377G 512K 377G 1% /dev/shm tmpfs 377G 1.8G 376G 1% /run tmpfs 377G 0 377G 0% /sys/fs/cgroup /dev/mapper/rhel_mathhpavw1202a-root 7.0T 5.4T 1.7T 77% /
Quota management
If you run out of quota in your home space on the file server, leverage the local space described above.
Check your quota. The -human flag prints it in an easier-to-read format:
fs lq -human
Check what is using the quota:
du -h ~<username> > /tmp/<username>.usage
Then look at the /tmp/<username>.usage file to see where the storage is being used. It's typically in .cache or other dot-file locations, and these can be moved to /export/<username>.
Note that /tmp is temporary space that can and will be cleared out by the operating system on reboots or when needed. Don't use it for anything important. The size of this partition is also small relative to the other local storage mounted on /export.
Redirecting large directories to /export
For example, if you are trying to install something like miniconda3, you can create a symbolic link so that anything written to ~<username>/miniconda3 is redirected to /export/<username>/miniconda3:
ln -s /export/<username>/miniconda3 ~/miniconda3
To move .cache, so all future data written to ~/.cache is redirected to /export/<username>/.cache:
mv ~/.cache /export/<username> ln -s /export/<username>/.cache ~/.cache
See Issues below if you need assistance with this.
Long running jobs
- For jobs that you think will run over 24 hours, don't use your home space. Use the temporary
/export/<username>directory to store input and output files — the file server where your home space is located expires access every 24 hours for security reasons. - If you end your SSH session, anything you have running will terminate as well. There are multiple ways to handle this, but we recommend using
screenortmux:
Accessing software
If you run a command to execute a piece of software and the command isn't found, you can use the module command to see if the software is available that way.
Issues
Contact math-helpdesk@umd.edu, call 301-405-1634, or stop by Martin Hall 0123 for assistance, including requests for any software installations.
UMD HPC
Zaratan is available, but Math has not made any investments. There are some free allocation options available. See hpcc.umd.edu.