Problem: Powerful PC with RTX 3090 consumes hundreds of watts and cannot run for days from UPS during power outages
Powerful graphics cards like the NVIDIA RTX 3090 require a significant amount of energy to operate. In home environments where power supply can be unreliable, this becomes a serious issue. Using an Uninterruptible Power Supply (UPS) allows the system to continue operating for a short period after the main power is cut off, but for extended operation from the battery, a powerful and expensive UPS is required.
This article discusses the architecture of a home AI server that includes automatic task management and switching to cloud resources when necessary. This ensures continuous operation of automation (n8n, bots, webhooks, website) 24/7, even during prolonged power outages.
Automation and Critical Tasks
Process automation using tools like n8n, bots, and webhooks requires constant access to resources. In the event of a power outage, such systems must be able to switch to backup power sources or cloud services without losing performance.
To solve this problem, an architecture was developed that includes Control Plane and Worker Plane.
Architecture: Control Plane vs Worker Plane
+---------------------+
| Control Plane |
| +-----------------+|
| | Ubuntu Mini ||
| | Server (wplanc)||-------> Docker, n8n, Nginx, MariaDB, WordPress, task queues
| +-----------------+|
| DC-UPS 12В |
+---------------------+
|
v
+---------------------+
| Router Keenetic |
+---------------------+
+---------------------+
| Worker Plane |
| +-----------------+|
| | Xeon LGA2011-v3 ||-------> LM Studio (port 1234), ComfyUI (port 8188)
| | RTX 3090 24GB ||
| +-----------------+|
+---------------------+
Control Plane: Low-power Ubuntu mini-server on DC-UPS 12V
The Control Plane includes a low-power Ubuntu mini-server running from a DC-UPS 12V. This server is responsible for task management and ensuring their continuous operation. It has Docker, n8n, Nginx, MariaDB, WordPress, and task queues installed.
Using Docker makes it easy to deploy and scale applications, providing isolation and ease of management. n8n is used for process automation, while Nginx and MariaDB handle web sites and databases.
Worker Plane: Desktop on Xeon LGA2011-v3 with RTX 3090
The Worker Plane includes a powerful desktop computer based on the Xeon LGA2011-v3 processor (X99 motherboard, 32 GB quad-channel RAM) and an RTX 3090 graphics card with 24 GB of video memory. This node is responsible for performing heavy computational tasks.
LM Studio (port 1234) and ComfyUI (port 8188) are installed on the Worker Plane, used for working with AI models and graphical tasks. These applications require a significant amount of computing resources, making them dependent on stable power supply.
Mechanism of Auto-Restart and Fault Tolerance
To ensure continuous operation of the system during power outages, an auto-restart and fault tolerance mechanism was implemented. It includes BIOS settings, Wake-on-LAN via Keenetic router, and task management logic in n8n.
BIOS Settings (Restore on AC Power Loss) + Wake-on-LAN
The «Restore on AC Power Loss» function was enabled in the Worker Plane’s BIOS. This allows the system to automatically reboot when power is restored after an outage. For additional security and ease of management, the Wake-on-LAN feature via Keenetic router is used.
Logic in n8n: health-check for local GPU
A logic was implemented in n8n for monitoring system status. It includes a regular health-check for the local GPU (RTX 3090). If power is restored and the node is alive, tasks are executed on the RTX 3090. In case of a power outage, heavy tasks are placed in a local SQLite/Redis queue or switched to the cloud Gemini API.
Operational Experience and Upgrade Thoughts
During the operation of this architecture, several key points were identified that should be considered for further system development.
128 GB ECC RAM for 70B models vs second RTX 3090
One possible upgrade is increasing the amount of RAM to 128 GB with ECC (Error-Correcting Code) support. This would allow using larger AI models, such as 70B, which require a significant amount of RAM for efficient operation.
An alternative option is installing a second RTX 3090. This would increase the system’s computational power and parallelize tasks between two graphics cards, potentially significantly improving performance in some scenarios.
In conclusion, this home AI server architecture with Control Plane and Worker Plane separation ensures continuous operation of automation and critical tasks even during prolonged power outages. Using a low-power Ubuntu mini-server and a powerful desktop computer, along with implementing auto-restart and fault tolerance mechanisms, makes the system reliable and scalable.
Добавить комментарий