Source: Marc Mercer (SRE Lead) — sre-iac repository + Bryan Lee (Engineering Director), Rev 2.0, 2026-03-22
Each server section is labeled with its current operational status. See Architecture Overview for the full context of Current vs Future state.
Server Inventory
| Unit | Hostname | IP | iLO IP | Status | Role |
|---|
| HP #1 | pmx-01 | 10.10.96.5 | 10.10.96.237 | ✅ ACTIVE | Proxmox — hosts all K8s VMs and services |
| HP #2 | (pending) | (off) | 10.10.96.235 | 🔵 PENDING | OpenStack controller/compute target |
| HP #3 | (pending) | (off) | 10.10.96.236 | 🔵 PENDING | OpenStack controller/compute target |
| HP #4 | (pending) | (off) | (offline) | 🟡 COMING SOON | Open Air GPU Farm (Proxmox, 3× GPU initially) |
| HP #5 | (planned) | (TBD) | (TBD) | 📋 PLANNED | OpenStack backup/spare node (not yet purchased) |
The only operational compute node. Hosts all current K8s VMs, application VMs, FreeIPA DCs, and the Proxmox hypervisor.
CPU
| Component | Specification | Per Server |
|---|
| Processor | Intel Xeon E5-2697 v4 | 2 CPUs |
| Architecture | Broadwell, 14nm | — |
| Base Clock | 2.3 GHz | — |
| Turbo Clock | 3.6 GHz | — |
| Cores per CPU | 18 cores | 36 cores total |
| Threads per CPU | 36 threads | 72 threads total |
| L3 Cache | 45 MB per CPU | 90 MB |
| TDP | 145W per CPU | 290W |
CPU features: VT-x, EPT, VT-d, AES-NI, Hyper-Threading, Turbo Boost 2.0.
Memory
| Component | Specification |
|---|
| Memory Type | DDR4-2133 ECC RDIMM |
| Module Size | 32GB |
| Total Memory | 512GB (16× 32GB) |
| Memory Channels | 8 per CPU (16 total) |
Storage
| Component | Specification | Purpose |
|---|
| OS Drive | 1TB USB SSD | Proxmox VE boot and OS |
| Ceph SSD (ready) | 8× 1TB enterprise SSD | Awaiting Ceph cluster deployment |
| Ceph HDD (ready) | 16× 1TB enterprise HDD | Awaiting Ceph cluster deployment |
| Storage Controller | HP Smart Array P440ar (HBA mode) | Direct disk access for Ceph |
Network Interfaces
| Component | Specification |
|---|
| Built-in NIC | Intel I350 Quad Port 1GbE — 4 ports |
| Bond: bond-net | eno1 + eno2 (LACP 802.3ad) — 2 Gbps aggregate |
| Bond: bond-stor | eno3 + eno4 (LACP 802.3ad) — 2 Gbps aggregate |
| Upgrade NIC | 10Gtek 10Gb Dual SFP+ (Intel 82599ES) — being installed |
| Management | Dedicated iLO 4 NIC (10.10.96.237) |
Power
| Component | Specification |
|---|
| PSU | 2× 800W redundant hot-plug |
| Cooling | Redundant hot-swap fans |
| Form Factor | 2U tower (rack-mountable) |
HP #2 — OpenStack Controller Target [PENDING — POWERED OFF]
Powered off. iLO accessible at 10.10.96.235. Hardware identical to HP #1. Will become an OpenStack controller/compute node when the OpenStack deployment phase begins.
Action required before deployment: Connect to iLO, verify BIOS, RAID controller in HBA mode, memory seating.
Specifications (identical to HP #1)
| Component | Specification |
|---|
| Processor | 2× Intel Xeon E5-2697 v4 (36 cores / 72 threads) |
| Memory | 512GB DDR4-2133 ECC RDIMM |
| SSD Storage | 8× 1TB enterprise SSD (Ceph SSD pool) |
| HDD Storage | 16× 1TB enterprise HDD (Ceph HDD pool) |
| Built-in NIC | Intel I350 Quad Port 1GbE |
| Upgrade NIC | 10Gtek 10Gb Dual SFP+ (Intel 82599ES) — to be installed |
| Management | iLO 4 (10.10.96.235) |
HP #3 — OpenStack Compute Target [PENDING — POWERED OFF]
Powered off. iLO accessible at 10.10.96.236. Hardware identical to HP #1 and HP #2. Will become an OpenStack compute node in the same deployment phase as HP #2.
Action required before deployment: Connect to iLO (10.10.96.236), verify BIOS, RAID controller mode, memory seating.
Specifications (identical to HP #1 and HP #2)
| Component | Specification |
|---|
| Processor | 2× Intel Xeon E5-2697 v4 (36 cores / 72 threads) |
| Memory | 512GB DDR4-2133 ECC RDIMM |
| SSD Storage | 8× 1TB enterprise SSD (Ceph SSD pool) |
| HDD Storage | 16× 1TB enterprise HDD (Ceph HDD pool) |
| Built-in NIC | Intel I350 Quad Port 1GbE |
| Upgrade NIC | 10Gtek 10Gb Dual SFP+ (Intel 82599ES) — to be installed |
| Management | iLO 4 (10.10.96.236) |
HP #4 — Open Air GPU Farm [COMING SOON — POWERED OFF]
HP #4 is an HP ProLiant ML350 Gen9 that has been stripped of its outer chassis enclosure and remounted in an open-air rack frame with 12 active cooling fans. The ML350 motherboard, CPUs, RAM, PSUs, and cabling remain unchanged — only the tower enclosure has been removed. This configuration allows NVIDIA GPUs to be mounted externally via PCIe risers, accommodating larger GPU form factors that do not fit inside the standard tower.
This node runs Proxmox VE independently and is not part of the OpenStack cluster. No GPUs reside in any OpenStack node.
CPU and Memory
| Component | Specification |
|---|
| Processor | 2× Intel Xeon E5-2683 v4 (32 cores / 64 threads) |
| Base / Turbo Clock | 2.1 GHz / 3.0 GHz |
| L3 Cache | 40 MB per CPU (80 MB total) |
| TDP | 120W per CPU (240W total) |
| Memory | 128GB DDR4 ECC RDIMM |
OS Storage
| Component | Specification | Purpose |
|---|
| Primary OS | 1TB USB SSD | Proxmox VE OS, VM disk images |
| Recovery | 64GB USB thumbdrive | Proxmox recovery ISO, emergency boot |
No internal drive array. No Ceph OSDs. No contribution to storage cluster.
GPU Configuration
| Slot | GPU | Architecture | VRAM | Memory Bus | TDP | Status |
|---|
| 1 | NVIDIA RTX 8000 | Turing (TU102) | 48GB GDDR6 | 384-bit | 260W | ✅ Installed |
| 2 | NVIDIA RTX 3090 | Ampere (GA102) | 24GB GDDR6X | 384-bit | 350W | ✅ Installed |
| 3 | NVIDIA RTX 3090 | Ampere (GA102) | 24GB GDDR6X | 384-bit | 350W | ✅ Installed |
| 4 | TBD | — | — | — | — | 📋 Reserved — will be selected post go-live |
Total VRAM (initial): 96GB (48 + 24 + 24)
The 4th GPU will be selected based on observed workload demand after the farm goes live. Likely candidates: NVIDIA RTX 4090 (24GB), NVIDIA A5000 (24GB), or a second RTX 8000 (48GB) if large-model demand is high.
GPU Passthrough — Proxmox VM Layout
Each GPU is assigned to a dedicated Proxmox VM via PCIe passthrough (VT-d). This provides near-native GPU performance and complete isolation between workloads.
| VM | GPU | vCPU | RAM | Primary Services |
|---|
| gpu-vm-01 | RTX 8000 (48GB) | 16 | 48GB | Ollama (large models 70B+), vLLM |
| gpu-vm-02 | RTX 3090 #1 (24GB) | 8 | 40GB | Ollama (mid-size), Faster-Whisper ASR, TTS |
| gpu-vm-03 | RTX 3090 #2 (24GB) | 8 | 40GB | Fine-tuning (Axolotl/Unsloth), batch inference |
| gpu-vm-04 | Reserved (4th GPU) | TBD | TBD | TBD post go-live |
See Architecture Overview — GPU Farm for full VM architecture detail and AI services stack.
Network Interfaces
| Component | Specification |
|---|
| NIC | 10Gtek 10Gb Dual SFP+ (Intel 82599ES) — to be installed |
| Port 1 (bond-net) | → TL-SX3008F-01 (network plane) |
| Port 2 (bond-stor) | → TL-SX3008F-02 (storage plane — NFS to QNAP / future Ceph) |
| Management | iLO 4 (offline — reconnect before power-on) |
Cooling
12 active fans in open-air rack frame. GPU TDP budget: RTX 8000 (260W) + 2× RTX 3090 (350W each) = 960W GPU-only. Combined with CPUs (~240W) and RAM, total draw approaches ~1,300–1,400W under full load. Verify PSU capacity and UPS assignment before powering on.
HP #5 — OpenStack Spare/Backup Node [PLANNED — NOT PURCHASED]
A fifth HP ProLiant ML350 Gen9 with specifications identical to HP #1–#3 will be purchased once the OpenStack cluster is fully live and stable. This node serves as:
- OpenStack spare: Standby node for rolling maintenance (patch one node while keeping full capacity)
- Capacity buffer: Absorbs workload if any of HP #2 or HP #3 fails unexpectedly
- Future expansion: If workload demands a fourth compute node, HP #5 is ready to join
| Specification | Value (planned) |
|---|
| CPU | 2× Intel Xeon E5-2697 v4 |
| Memory | 512GB DDR4-2133 ECC RDIMM |
| SSD | 8× 1TB enterprise SSD (Ceph SSD pool) |
| HDD | 16× 1TB enterprise HDD (Ceph HDD pool) |
| NIC | Intel I350 Quad Port 1GbE + 10Gtek dual SFP+ |
10G NIC Upgrade — All ML350 Servers
Four 10Gtek 10Gb PCI-E NIC cards (Intel 82599ES chipset, dual SFP+ ports) have been ordered — one per ML350 server (HP #1, #2, #3, #4). This NIC provides two independent 10G SFP+ ports, one per switch plane:
| Server | Port 1 (bond-net) | Port 2 (bond-stor) |
|---|
| HP #1 (pmx-01) | → SX3008F-01 (network 10G) | → SX3008F-02 (storage 10G, isolated) |
| HP #2 (OpenStack) | → SX3008F-01 | → SX3008F-02 |
| HP #3 (OpenStack) | → SX3008F-01 | → SX3008F-02 |
| HP #4 (GPU Farm) | → SX3008F-01 | → SX3008F-02 |
DAC cables ordered: 8× H!Fiber 1.5m SFP+ DAC (4 servers × 2 planes). Additional 0.3m DACs needed for switch-to-router and cross-panel connections.
Aggregate Totals
Current State (HP #1 only)
| Resource | Value |
|---|
| CPU Cores | 36 cores / 72 threads |
| RAM | 512GB |
| Proxmox VMs | 8+ VMs (6 K8s nodes + application VMs) |
| Storage (NFS) | QNAP qnap-01 (10.10.96.31) — NFS for all K8s PVs |
Future State — OpenStack Cluster (HP #1 + #2 + #3)
| Resource | Total |
|---|
| CPU Cores | 108 cores / 216 threads |
| RAM | 1,536GB (1.5TB) |
| SSD (Ceph raw) | 24TB (8 per node × 3 nodes) |
| SSD (Ceph usable, RF=2) | 12TB |
| HDD (Ceph raw) | 48TB (16 per node × 3 nodes) |
| HDD (Ceph usable, RF=2) | 24TB |
Future State — GPU Farm (HP #4, initial 3 GPUs)
| Resource | Value |
|---|
| CPU Cores | 32 cores / 64 threads |
| RAM | 128GB |
| VRAM | 96GB (48 + 24 + 24) |
| GPU Slots | 3 active / 1 reserved |
Combined Future State (HP #1–#4, excluding HP #5)
| Resource | Total |
|---|
| CPU Cores | 140 cores / 280 threads |
| RAM | 1,664GB |
| VRAM | 96GB (expanding to TBD with 4th GPU) |
| Raw SSD | 24TB |
| Raw HDD | 48TB |
Network Hardware
Current State (Juniper — Unmanaged)
| Device | Model | IP | Status | Role |
|---|
| Cisco Router | (unknown model) | 10.10.96.1 | ✅ Active | Internet gateway, bridges lab + homestead (flat /20) |
| EX2200-24T #1 | Juniper EX2200-24T | (unmanaged) | ✅ Active, unmanaged | L2 switch — no VLANs |
| EX2200-24T #2 | Juniper EX2200-24T | (unmanaged) | ✅ Active, unmanaged | L2 switch — no VLANs |
| EX2200-24P #1 | Juniper EX2200-24P | (unmanaged) | ✅ Active, unmanaged | L2 switch — no VLANs |
| QNAP NAS #1 | QNAP (model TBD) | 10.10.96.31 | ✅ Active | NFS storage for K8s PVs |
| QNAP NAS #2 | QNAP (model TBD) | 10.10.101.142 | ❌ Offline (bad drive) | Decommissioned until repaired |
Interim Configuration (Juniper — VLANs Pending)
| Device | Hostname | Model | Ports | Role |
|---|
| Edge Firewall | srx320-01 | Juniper SRX320 | 6× GbE + 2× SFP | WAN, inter-VLAN routing, security zones |
| Network Switch | ex2200t-01 | Juniper EX2200-24T | 24× GbE + 4× SFP | Network plane (VLANs 200, 500, 600, 610, 1000+) |
| Storage Switch | ex2200t-02 | Juniper EX2200-24T | 24× GbE + 4× SFP | Storage plane (VLANs 300, 400) — isolated |
| OOB Switch | ex2200p-01 | Juniper EX2200-24P | 24× GbE PoE + 4× SFP | Management plane (VLAN 100) — isolated |
SRX320 config is deployment-ready in sre-iac/network/interim/srx320-01.conf. EX2200 VLAN configs are in sre-iac/network/interim/.
Target Configuration (OPNsense Edge + TP-Link Omada SDN switches + Juniper OOB)
| Device | Hostname | Model | Ports | Role |
|---|
| Firewall Primary | opnsense-01 | Beelink EQ12 Pro + OPNsense CE | 2× 2.5GbE Intel i226-V | WAN edge, IDS/IPS, VPN, CARP primary, HAProxy LB |
| Firewall Standby | opnsense-02 | Beelink EQ12 Pro + OPNsense CE | 2× 2.5GbE Intel i226-V | CARP standby — auto-failover if primary fails |
| 10G Network Switch | sx3008-01 | TP-Link TL-SX3008F | 8× 10G SFP+ | Network plane (server bond-net uplinks + OPNsense LAN) |
| 10G Storage Switch | sx3008-02 | TP-Link TL-SX3008F | 8× 10G SFP+ | Storage plane (server bond-stor + QNAP) — isolated, no WAN uplink |
| 1G Network Switch | sg3428-01 | TP-Link TL-SG3428 | 24× GbE + 4× SFP | Network plane 1G access (control VMs, workstations) |
| 1G Storage Switch | sg3428-02 | TP-Link TL-SG3428 | 24× GbE + 4× SFP | Storage plane 1G fallback — isolated |
| OOB Switch | ex2200p-01 | Juniper EX2200-24P | 24× GbE PoE + 4× SFP | Management plane (VLAN 100) — repurposed from interim |
| Homestead Switch | sg3452-01 | TP-Link TL-SG3452 | 48× GbE + 4× SFP | Residential network (isolated from lab) |
| SDN Controller | oc200-01 | TP-Link Omada OC200 | 1× GbE | Omada SDN — manages all TP-Link switches (PoE from OOB switch) |
There is no TP-Link ER-series router in this design. OPNsense on the Beelink EQ12 Pro pair handles all WAN, firewall, IDS/IPS, VPN, and inter-VLAN routing. The Omada SDN controller manages only the switch fabric. This saves ~$399 (ER7412-M2 cost) and delivers a far superior security posture.
See Network Edge Security for the full OPNsense architecture.
Other Devices
| Device | Hostname | IP | Connection | Role |
|---|
| BeeLink Mini PC | blnk-fed4201 | 10.10.96.3 | 2× 1G (one per EX2200) | Dev workstation, scan host, management access |
| QNAP NAS #1 | qnap-01 | 10.10.96.31 | 1G to switch | NFS storage for all K8s PVs |
| QNAP NAS #2 | (unnamed) | 10.10.101.142 | offline | Bad drive — treat as decommissioned |
| UPS #1 | smt15c-apc01 | (TBD) | 1G to OOB | Power management |
| UPS #2 | smt15c-apc02 | (TBD) | 1G to OOB | Power management |
| UPS #3 | smt15c-apc03 | (TBD) | 1G to OOB | Power management |
| UPS #4 | smt15c-apc04 | (TBD) | 1G to OOB | Power management |
Physical Specifications — HP ProLiant ML350 Gen9
| Measurement | Specification |
|---|
| Form Factor | 2U Tower (rack-mountable) |
| Height | 174.5 mm (6.87 in) |
| Width | 442.0 mm (17.4 in) |
| Depth | 660.0 mm (26.0 in) |
| Weight (loaded) | ~35 kg (77 lbs) |
| Operating Temperature | 10°C to 35°C |
| Humidity | 8% to 90% RH (non-condensing) |
Note (HP #4 only): Physical form factor does not apply — outer chassis removed. Motherboard, PSUs, and cabling are mounted in an open-air rack frame with 12 fans. Environmental tolerances remain the same per component specs.
Document Control
| Rev | Date | Author | Description |
|---|
| 1.0 | 2026-02-24 | Marc Mercer | Initial release |
| 2.0 | 2026-03-22 | Anshin Engineering | Added server status labels (active/pending/planned); corrected HP #4 as open-air ML350 (not standard tower); added HP #5 planned node; added 10G NIC upgrade specs; added GPU farm VM layout; corrected router spec (ER7412-M2 not ER8411); added current state network table |