Slingshot Fabric Manager on baremetal

Introduction

The Slingshot Fabric Manager (FM) on bare-metal enablement within the Cray System Management (CSM) framework introduces dedicated Fabric Manager Nodes (FMNs) that manage and monitor Slingshot fabric operations outside of the Kubernetes environment.

CSM 1.7.1 includes bare-metal FM support, which provides the necessary base OS image, networking, and storage configurations for running the Slingshot Fabric Manager natively within the CSM environment.

Please note that this feature is only available from CSM 1.7.1 onwards.

NOTE:

  • FMNs are considered Management nodes.
  • After Fabric Manager is migrated from a Kubernetes pod to bare-metal infrastructure, it cannot be reverted.
  • The two FMNs must be part of two different management racks to support Rack Resiliency.
  • This feature will not be supported on systems with Dell/ Mellanox based management networks.

Terminology and Components

Component Reference
SHS Slingshot Host Software
FM Fabric Manager
FMN Fabric Manager Node
SLS System Layout Service
HSM Hardware State Manager
BSS Boot Script Service
CANU CSM Automatic Network Utility
SAT System Admin Toolkit
SMA System Monitoring Application

Architecture

In CSM versions <= 1.7.0, the deployment of the Fabric Manager within CSM uses native Kubernetes capabilities—both during upgrades and in failure/HA scenarios. Kubernetes itself provides health checks and a scheduler that can rebalance workloads across nodes based on load, administrative policies, and other criteria. The Fabric Manager is deployed as a single pod in Kubernetes. A traditional HA model for Fabric Manager doesn’t map cleanly into Kubernetes, so instead, Kubernetes’ built-in mechanisms detect failures and spin up a replacement pod, minimizing downtime.

In theory, this model should satisfy HA requirements: if the pod fails (or needs to be moved during an upgrade), Kubernetes can detect the fault and recreate the Fabric Manager on another node, providing continuity.

In practice, however, this approach does not meet the contractual HA obligations. Because of Kubernetes “best‑effort” scheduling and the resource demands of the Fabric Manager, real service outages can exceed 5 minutes.

To address these issues, CSM 1.7.1 includes FM on baremetal support, which provides the necessary base OS image, networking, and storage configurations for running the Slingshot Fabric Manager natively within the CSM environment to achieve HA.

Configure FM on baremetal

To configure FM on baremetal please follow the procedure.

Slingshot Switch Firmware Update

  • For clusters using the FM pod: CSM will continue to handle switch firmware uploads and updates as before.
  • For clusters with bare-metal FM: FMN will host the switch firmware, and FM will be responsible for managing switch updates. Refer section “3.2.6 (Optional) Update HPE Slingshot switch firmware” in HPE Slingshot Installation Guide for CSM PDF.