This project documents the shared host built for the DanSTAR aerospace team. The job was to replace fragmented cloud services and personal workstations with one box for CAD, data analysis, version control, and file storage.
Hardware and expansion
The server is an AMD Ryzen Threadripper 9960X on a TRX50 platform with 128GB of ECC DDR5.
We evaluated AM5 first, then moved to Threadripper for PCIe lanes, memory bandwidth, and expansion. CAD needs strong single-core performance and large memory footprints. ECC was a hard requirement: long-running CI runners and databases cannot silently corrupt.
We chose tower air cooling over liquid. A dead pump shuts the machine down at once. A dead fan raises temperature slowly enough to react.
Virtualization and service boundaries
The system runs Proxmox VE. Workloads are isolated on purpose:
- LXC Containers: Lightweight Linux containers run Docker Engine. We dropped Portainer for plain Docker Compose. A container web UI adds state we did not want. Version-controlled
docker-compose.ymlfiles are reproducible and readable. GitLab runs in its own LXC for repositories and CI/CD. - Network Segmentation: Virtual networks isolate guest workloads from the main LAN.
iptablesmasquerading gives outbound internet without exposing internal services.
Storage and persistence
TrueNAS Core runs in a VM with PCIe passthrough of the mechanical drives. The disks form a 24TB RAIDZ1 ZFS pool.
That pool serves NFS and Samba for team files and container storage. Automated ZFS snapshots run weekly. Off-site backups sync to cloud storage for a 3-2-1 backup rule.
Development environments and GPU passthrough
We provisioned dedicated development VMs for mechanical engineering.
The primary CAD environment is a Windows 11 Pro VM. We used tiny11 scripts to strip telemetry, background services, and unused apps. Idle memory use fell by half. Baseline responsiveness improved.
For Ansys hardware acceleration we passed a GTX 970 into the Windows VM with IOMMU grouping. The Proxmox host uses the motherboard’s integrated graphics. The discrete GPU stays bound to the Windows guest with native NVIDIA drivers.
A separate Ubuntu Server VM is an SSH sandbox for compilation and Linux-native work.
Why the host exists
This is shared infrastructure, not a hardware showcase. Software and operations stay on separate sides: allocated resources, versioned configs, hardware limits treated as software limits.
Leadership context
I drove the move from fragmented cloud services and personal workstations to one platform for the DanSTAR software team. Shared decisions: Proxmox workload isolation, versioned Docker Compose instead of ad hoc container UIs, ZFS snapshots with off-site backup, and GPU passthrough VMs where CAD needs native tooling. The Parhelion program hub links this host to flight software, OS integration, and operator procedures.