1 of 10

2 of 10

Inband Management Agent

George Kola (Google) 🔉

Murugasamy Nachimuthu (Intel), Panos Christeas (Meta), Changho Choi (Samsung), Chukwunenye Nnebe (Microsoft)

3 of 10

Inband Management Agent

Definition

  • Piece of software that runs on host CPU and enables system management

Alternative

  • Out-of-band BMC based solution

​

Use Cases

  • monitoring
  • telemetry
  • firmware-update
  • failure prediction
  • self-healing
    • memory poison containment
    • DIMM Post-Package-Repair

​

4 of 10

Why inband ?

Certain monitoring/management is available only inband

  • Host kernel/software metrics
  • Host CPU metrics
  • PCIE error metrics
  • Memory Check Exception (MCE)
  • Memory poison containment
  • Memory Post Package Repair

​

Direct high bandwidth connection to PCIE

  • Firmware updates (GPU card, PCIE cards)

​

Higher processing power of server CPU

​

Augment Out-of-Band

  • Machines without BMC
  • Machines with limited microcontroller
  • Inband while server CPU works and out-of-band on CPU failure

​

Inband in management server/processor

  • Rack management unit
  • Switch Management processor
  • Smart NIC

​

​

5 of 10

Inband Architecture

Agents per device class

Coordinator has list of available agents

  • Activates agents based on devices found (dynamic) or pre-configured (static)
  • Agents export telemetry that is consolidated into Redfish tree by co-ordinator
  • Coordinator provides common functionality and access control. Time lease for access.
  • Management container/Host OS ensure safe access

​

​

Device Agent 1

Device Agent 2

Device Agent N

Coordinator

RedFish + OEM OCP IMA extension Interface

Inband Management Agent (IMA)

Device Class 1

........

........

enclave

Device Class 1

Device Class 1

6 of 10

Agent Discovery

  • Each agent specifies the device class it supports
    • Can be inspected before enabling agent
  • Static or Dynamic activation
    • Static: Preconfigured list of agents to activate
      • Can be done based on devices on a server SKU
    • Dynamic: Coordinator activates agents based on hardware detected
      • Allows flexibility

7 of 10

Telemetry API (IMA Northbound API)

  • Coordinator will export DMTF Redfish tree with data collected across the agents
    • We will work with DMTF on additional modeling needed
  • Southbound API
    • Device hardware specific
    • We are not specifying a protocol.

8 of 10

Security Architecture

  • No direct access from agent to Internet/off host
  • Telemetry access provided via co-ordinator
    • Co-ordinator limits data exported
      • Prevents accidental PII leak
  • Hardware access enforced by Management container/Host OS
    • Sensitive access time limited (request, use, return)
    • E.g. reading DIMM limited to early host boot when there is no PII
    • Coordinator + core-library provide primitives to request access

9 of 10

Join Us !

10 of 10

Open Discussion