Announcing the Hopper Cluster (Hopper)
The ORC would like to invite you to use Hopper its new high performance compute cluster. Hopper is named in honor of the late rear admiral Grace Hopper, a computing pioneer and local resident. All new ORC cluster accounts will be created and activated on Hopper by default. However, existing Argo cluster account holders should send an email to orchelp@gmu.edu to request activation of their account on Hopper.
Hopper currently has a total of 70 compute nodes each node with 48 cores (Intel Cascade Lake) and 188 GB of available memory. Currently, 28 nodes and the GPU node are freely available for all users. The remaining nodes may also be used but jobs will be subject to preemption by jobs run by the node’s sponsors. There is one Nvidia DGX GPU node with 128 CPU cores (AMD EPYC/Milan), 1 TB of memory, and 8xA100 GPUs.
A large expansion of Hopper is planned for the Fall of 2021 which will add a substantial number of compute and GPU nodes including very large memory nodes with up to 4 TB of memory. Users who require memory address spaces greater that 180 GB will need to continue to use the Argo cluster until the new large memory nodes become available in Hopper.
The Hopper cluster is configured in a similar but not identical fashion to Argo. The software modules are organized differently and there are differences in the partition names, defaults, and versions of software available. Please review the documentation linked below for more detailed information on the differences.
You may log in to Hopper using “ssh <UserID>@hopper.orc.gmu.edu,” where “<UserID>“ is your GMU NetID, use your GMU campus password when prompted. Home, scratch, and project directories will be mounted in the same locations as on Argo. Let us know if there are any “groups” directories you need to access, or if there are specific software packages and versions you require that are not available. The partition/queue structure on Hopper is summarized in the table below:
| Partition | Time Limit (D-H:M) | Description | ARGO Equivalent |
| debug | 0-01:00 | Intended for quick tests | |
| interactive | 0-12:00 | Interactive jobs (Open OnDemand) | |
| normal | 3-00:00 | default | all-LoPri, all-HiPri, bigmem-HiPri, bigmem-LoPri, all-long, bigmem-long |
| contrib* | 6-00:00 | CDS_q, COS_q, CS_q, EMH_q | |
| gpuq | 1-00:00 | GPU node access | gpuq |
*NOTE: Being a contributor on Argo does not automatically grant access to the contrib partition on Hopper. All users may submit jobs to the contrib partition on Hopper, however, their jobs may be preempted and killed by a contributor’s job at any time. We recommend that non-contributor users who submit to the contrib partition ensure their jobs use some form of checkpointing. Contact orchelp@gmu.edu if you need help implementing checkpointing in your jobs.
Open OnDemand
We would also like users to try our new Open OnDemand (OOD) Server, which enables launching interactive apps including RStudio, Jupyter Lab, MATLAB and Mathematica, or a Linux graphical desktop through a web interface. These interactive sessions can be used for up to 12 hours. From a web browser, login to https://ondemand.orc.gmu.edu using your GMU username and credentials to access the OOD server. Please let us know of any problems you encounter, and any applications you would like to be able to use via Open OnDemand.
Documentation
Please refer to the following links for current documentation on Hopper:
- Getting Started: http://wiki.orc.gmu.edu/mkdocs/Hopper_Quick_Start_Guide/
- Navigating Modules on Hopper: http://wiki.orc.gmu.edu/mkdocs/Navigating_Lmod_Modules/
- Running GPU jobs on the DGX A100: http://wiki.orc.gmu.edu/mkdocs/DGX_A100_User_Guide/
- Open OnDemand on Hopper: http://wiki.orc.gmu.edu/mkdocs/open_ondemand_on_Hopper/
- Comparing ARGO to HOPPER: http://wiki.orc.gmu.edu/mkdocs/ARGO_vs_HOPPER/
If you have any questions about any aspect of the new Hopper cluster, please send an email to orchelp@gmu.edu.