Engineering the Hybrid Enterprise Homelab: 10G SAN MPIO, Local GPU LLMs & Hardened VLANs

Enterprise Systems Architecture

Engineering the Hybrid Enterprise Homelab

Building a resilient, production-grade systems laboratory combining bare-metal hypervisors, 10G iSCSI storage multipathing, container microservices, and private local GPU intelligence.

XCP-ng Dom0
Dell EqualLogic PS4100
iSCSI MPIO Round-Robin
NVIDIA RTX 4070 SUPER
UniFi UDM-Pro
Active Directory Split-DNS

1. The Vision: Enterprise Patterns in a Private Lab

Most homelabs start as a chaotic sprawl of single-board computers and consumer routers. When architecting customer solutions for IT Tech STL, testing requires the exact enterprise disciplines used in mid-market data centers: redundant storage fabrics, strict VLAN micro-segmentation, active directory split-DNS, and resilient virtualization clusters.

Here is how our infrastructure was engineered from physical copper up to autonomous container orchestration.

2. Dual-Path Storage Multipathing (Dell EqualLogic PS4100)

Storage bottlenecks kill hypervisor performance. Rather than using brittle SMB or single-stream NFS mounts, central block storage is anchored by an enterprise Dell EqualLogic PS4100 SAN operating on dedicated non-routable iSCSI VLANs (VLAN 98 & 99).

# Verified Multi-Path Status on XCP-ng Dom0:
368fc61b6ab2b834df8abe6580c8a31a7 dm-0 EQLOGIC, 100E-00 (800GB)
`-+- policy=’round-robin 0′ prio=1 status=active
  |- 4:0:0:0 sdb 8:16 active ready running [Portal: 10.1.99.25:3260]
  |- 6:0:0:0 sdd 8:48 active ready running [Portal: 10.1.99.25:3260]
  `- 5:0:0:0 sdc 8:32 active ready running [Portal: 10.1.99.26:3260]
Latency: 0.16ms – 0.18ms | Flow Control: 802.3x RX/TX Active
  • Zero-Routing Architecture: iSCSI VLANs have DHCP strictly disabled and are firewalled from all RFC1918 subnets, eliminating broadcast overhead.
  • Round-Robin Striping: IOPS are distributed across physical host adapters (eth2 & eth3) with 802.3x Ethernet pause frames enabled end-to-end to prevent buffer drops.

3. Private AI Core: 45+ Tokens/Sec Local LLM Inference

Enterprise engineering in 2026 demands AI agility without surrendering proprietary network configurations, customer credentials, or topology maps to external public clouds.

Our primary engineering workstation (Deathstar) features an NVIDIA GeForce RTX 4070 SUPER running our customized architecture model (ittechstl-architect:latest, 14B Qwen base).

Initial Baseline
15.4 tok/s
18% CPU / 82% GPU Split
(32k Over-Allocated Context)

Optimized State
45.2 tok/s
100% Pure GPU Residency
(16k Ctx + Q8 KV Cache)

Sustained Speedup
+293%
Near-Zero PCIe Bus Latency
Flash Attention Active

4. Network Topography & VLAN Micro-Segmentation

Anchored by a UniFi Dream Machine Pro and dual 24-port PoE switches, our network enforces strict isolation boundaries:

VLAN Subnet Purpose Firewall Enforcement
VLAN 1 10.1.1.0/24 Management Dedicated out-of-band switch & AP access
VLAN 10 10.1.10.0/24 Workstations Authoritative DNS mapped to DC01
VLAN 20 10.1.20.0/24 Core Servers Inter-VLAN stateful inspection enabled
VLAN 40 10.1.40.0/24 IoT & Smart Home Client isolation & cloud-only egress
VLAN 98/99 10.1.98/99.0/24 iSCSI Storage Fabric Non-routable, zero RFC1918 cross-traffic

Next Architectural Milestone: High-Availability Pool

With storage multipathing verified, dev stacks containerized via Dockge, and local AI throughput stabilized at 45+ tok/s, our next milestone focuses on promoting our secondary Domain Controller (DC-AD2-01 on XCP-ng), transferring FSMO master roles, and converting our secondary Windows host (HV2) into our second bare-metal XCP-ng node to form a fully automated, high-availability hypervisor pool.

Leave a Comment