Skip to content

feat: Support hosts without DPUs #870

@Matthias247

Description

@Matthias247

NICo needs to support machines with potentially no DPUs in them. This is to be able to expand the possible deployments of NICo software to existing environments where there are no DPUs or where DPU installation is untenable for whatever reason (power, mechanical incompatibility, etc).

This will be taken up in multiple phases.

Hardware that does not support DPUs will not support:

  • Ethernet Network partitioning
  • Network Security Groups-
  • Transparent, fault-tolerant ethernet north/south connectivity via BGP ECMP
  • Metadata Service (FMDS)
  • Phone Home capability
  • Bring-your-own-IP
  • Routing table separation between tenants (Forge Native Networking)
  • Monitoring of the host from the DPU
  • Full mesh network connectivity / latency testing
  • ACLs to protect infrastructure (right now DPUs drop traffic for OOB networks, control plane, etc)

Zero-DPU machines will:

  • Have normal Redfish APIs that show up on the OOB network
  • Will be connected to a underlay network that's not controllable by NICo
  • Co-exist on the same site controller with single and multi-dpu hosts
  • Will not have a DPU (we need to distinguish between a machine that does not have a DPU vs one where the DPU is broken or Redfish can't see it)
  • NICo will not know the network the host is on until the host tries to DHCP for the first time (only here can we associate the DHCP client via MAC address to the Redfish NetworkAdapter data) - and know what its IP is.
  • be able to use network virtualization technologies that do not rely on DPUs (InfiniBand, NVLink)

Definition of Done:

A host with no DPUs in it (that still satisfies all of our redfish requirements) is able to:

  • Be ingested as a managed host
  • Can get used as an instance by a tenant
  • Skips DPU related things (including DPU upgrades)
  • Still can be monitored via Redfish & Hardware Health
  • Measured boot still works
  • SSH console still works
  • Tenant can still use all other non-DPU based features (e.g. IB and NVLink partitioning)

Metadata

Metadata

Assignees

Labels

featureFeature (deprecated - use issue type, but it's needed for reporting now)roadmapRoadmap item with program-level tracking

Type

No fields configured for Epic.

Projects

Status

In Progress

Milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions