Skip to main content

Bare Metal Provisioning

Business Value: Reduce GPU server provisioning from days of manual work to minutes of automated orchestration. No SSH, no manual OS installs, no network cable tracing.

How It Works​

The provisioning engine coordinates multiple components in a fully automated pipeline. When a tenant requests GPU nodes through the portal, the platform executes the following sequence without human intervention:

  • The network fabric controller prepares isolated VRFs and VLANs for the tenant
  • The bare metal controller powers on the physical server via IPMI/BMC, boots it over PXE, and installs a performance-optimized golden OS image
  • Cloud-init configures the node's hostname, networking, and deploys the GPU agent
  • The storage layer creates a dedicated parallel filesystem directory with InfiniBand access control
  • The GPU agent registers with the portal, and the node is ready for cluster assignment

Bare Metal Provisioning Flow

Technical Highlights​

  • PXE-based network boot with DHCP relay across provisioning VLANs
  • Golden OS image with pre-tuned BIOS settings: performance profile, C-states disabled, NUMA alignment enabled, PCIe ASPM disabled, virtualization disabled
  • OS-level tuning: CPU governor locked to performance, huge pages enabled, NUMA balancing disabled
  • Agent-based architecture: lightweight gRPC agent on every node for secure, bidirectional communication with the portal
  • Full deallocation flow: workload drain → network removal → storage cleanup → node returned to pool

Supported Hardware​

Enterprise GPU server platforms from multiple vendors, with configurable server profiles, GPU types, and accelerator configurations. Scalable to thousands of nodes per deployment.