About Hopper
The Hopper Cluster is a batch computing resource available to all faculty and their students at George Mason University.
Hardware
| Node Type | # Nodes | CPU Architecture | # Cores | Memory | GPUs |
|---|---|---|---|---|---|
| Login Nodes | 2 | Intel | 48 / node | 384 GB | No GPU |
| 2 | AMD | 64 / node | 256 GB | 1 Nvidia T4 | |
| Compute Nodes | 74 | Intel - 3552 cores | 48 / node | 192 GB | No GPU |
| 48 | AMD - 3072 cores | 64 / node | 256 GB | No GPU | |
| 20 | AMD - 1280 cores | 64 / node | 512 GB | No GPU | |
| 12 | AMD - 768 cores | 64 / node | 1024 GB | No GPU | |
| 8 | AMD - 1024 cores | 128 / node | 2048 GB | No GPU | |
| 2 | AMD - 256 cores | 128 / node | 4096 GB | No GPU | |
| GPU Nodes (A100) | 31 | AMD - 1984 cores | 64 / node | 512 GB | 124 Nvidia A100 - 80GB GPUs |
| GPU Nodes (DGX) | 2 | AMD - 256 cores | 128 / node | 1024 GB | 16 Nvidia A100 - 40GB GPUs |
| GPU Nodes (H100) | 1 | Intel - 112 cores | 112 / node | 2048 GB | 4 Nvidia H100 - 80GB GPUs |
Storage
/home- is subject to the same 60GB/user quota restriction as before/projects- is shared among members of a group or project/scratch- is an 1.5 PB (1500 TB) VAST flash-based high performance filesystem. Users' scratch files are subject to a 90-day purge policy
Heavy users can use Globus to transfer data to and from the /home, /projects and /scratch
Networking
- 100Gbps internal Mellanox HDR100 InfiniBand interconnect
- 25Gbps internal 25G ethernet network
- 20Gbps public connection via login nodes (20Gbps)
Software
Hopper is built using standard OpenHPC 2.x tools, namely
- Warewulf for provisioning custom CentOS 8 images to nodes
- Slurm for job scheduling and resource management
- OpenHPC repositories for essential HPC software
- Lmod modules for software provisioning to users
These applications are provisioned to users using Lmod modules.
Spack
Spack is used to build a lot of the software on top of the tools provided by OpenHPC
Containers
Users are encouraged to use Singularity containers to make applications more portable and easier to run. Users would need to convert Docker containers to Singularity before running them on Hopper.