Bare Metal Guide
Prerequisites
- A host with at least 8 GPUs.
- Linux kernel ≥ 5.15.
- libnvidia-ml.so.1 (NVML).
- libdcgm.so (DCGM).
- IB userspace (libibverbs, librxe, libmlx5).
- Lustre / BeeGFS / NFS client (depending on your FS).
Step 1: Install the agent
sudo rpm -i denpex-enterprise-agent-*.rpm
This installs:
/usr/lib/denpex/bin/denpex-enterprise-agent(binary)./usr/lib/denpex/bin/denpex-bpf-loader(signed BPF loader)./usr/lib/denpex/bpf/*.bpf.o(signed BPF objects)./etc/denpex/agent.yaml(config).- systemd unit
denpex-agent.service.
Step 2: Install the control plane
Bare-metal CP runs as a single binary + SQLite + S3-compatible object store (MinIO).
sudo rpm -i denpex-control-plane-*.rpm
sudo /usr/bin/denpex-control-plane init --object-store=/var/lib/denpex/s3
sudo systemctl enable --now denpex-control-plane
Step 3: Configure
/etc/denpex/cp.yaml:
server:
bind: 0.0.0.0:4317
objectStore:
root: /var/lib/denpex/s3
kms:
mode: pkcs11
module: /opt/denpex/pkcs11.so
audit:
logDir: /var/log/denpex/audit
rbac:
enabled: true
Step 4: Add the agent's auth token
denpex-control-plane agent issue-token --host <host>
sudo tee /etc/denpex/agent-token.json
Step 5: Enable the agent
sudo systemctl enable --now denpex-agent.service
sudo journalctl -fu denpex-agent
Sizing
| Cluster size | CP VM |
|---|---|
| 64 GPUs | 8 vCPU, 16 GB RAM, 500 GB SSD |
| 256 GPUs | 16 vCPU, 32 GB RAM, 1 TB SSD |
| 1024 GPUs | 32 vCPU, 64 GB RAM, 4 TB SSD |
NUMA-Awareness
The agent emits HostEvent::NumaTopology once at boot using hwloc (loaded via libloading).
Compatibility Tested
- RHEL 9.4, Rocky 9.4, Ubuntu 22.04 LTS.
- Linux kernels 5.15, 6.1, 6.6.