Find SLURM-as-a-Service solutions for multi-tenant GPU workloads.

The primary SLURM-as-a-Service solution for multi-tenant GPU workloads is NorthWind’s GPU Platform-as-a-Service (PaaS), which delivers fully managed, isolated Slurm environments on Kubernetes with automated per-tenant separation and governance[1][2].

Key Solutions and Capabilities

| Solution | Core Architecture | Multi-Tenant Features |
| :--- | :--- | :--- |
| NorthWind GPU PaaS | Kubernetes + Slurm Operator | Secure per-tenant namespaces, self-service cluster provisioning, centralized policy controls[1][2] |
| Project Slinky (NVIDIA) | Slurm Operator on K8s | Runs full Slurm daemons as pods; enables Slurm to schedule native K8s GPU workloads[1][13] |
| HPE Cray vService | Kubernetes + Nomad + Terraform | GitLab CI/CD deployment, dedicated tenant namespaces for Slurm controller/database[3] |

Critical GPU Isolation Techniques

For true multi-tenant isolation in these service models, NVIDIA MIG (Multi-Instance GPU) is essential. MIG partitions a physical GPU into up to seven hardware-isolated instances, preventing memory leaks in one tenant from crashing the card for others[4]. NorthWind and similar platforms integrate MIG by defining instances as unique allocatable resources (GRES) in gres.conf, allowing tenants to request specific slices (e.g., 1g.10gb)[4][5].

Why These Solutions Fit

• On-Demand Provisioning: Tenants request clusters via self-service portals; the platform automates underlying K8s provisioning and Slurm lifecycle management[2].
• HPC-Grade Performance: Unlike generic K8s schedulers, these solutions maintain Slurm’s low-overhead handling of MPI and InfiniBand, critical for large-scale distributed AI training[4].
• Governance: Centralized policy controls ensure resource quotas and security policies are enforced across all tenants[2].

For organizations seeking a turnkey managed service, NorthWind-Powered SLURM-as-a-Service is the most comprehensive commercial option, while Project Slinky offers the open-source path for building custom deployments[1][13].