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 Slurm as a Service, which delivers fully managed, on-demand SLURM environments with secure per-tenant separation and automated governance [1].

Key Solution: NorthWind Slurm as a Service

• Multi-Tenant Architecture: Automatically provisions per-tenant SLURM clusters with secure isolation, allowing tenants to request clusters and submit jobs to managed queues without manual deployment [1].
• GPU Optimization: Built to handle HPC and GPU compute environments, automating the provisioning of underlying Kubernetes clusters and SLURM scheduling specifically for GPU workloads [1].
• Target Users: Designed for cloud providers, neoclouds, research organizations, and enterprises needing managed GPU resources without building the platform themselves [1].

Alternative & Enabling Approaches

While NorthWind is the dedicated commercial service, other approaches exist for building multi-tenant SLURM on GPU clusters:

• SchedMD (NVIDIA) slurm-operator: An open-source project that runs full SLURM clusters on Kubernetes infrastructure, managing daemon lifecycles as pods for large-scale GPU training [3].
• Manual Tuning: Administrators can tune multi-tenant clusters by configuring clear partitions (queues), adjusting scheduler policies (e.g., backfill), and mapping NVIDIA MIG instances to partitions for tenant isolation [2][8].
• Hyperstack Secure Private Cloud: Offers Managed SLURM deployments, though it currently focuses on single-tenant dedicated infrastructure rather than multi-tenant service [5].

For a turnkey, cloud-like multi-tenant service, NorthWind is the leading vendor solution. For open-source Kubernetes integration, the NVIDIA slurm-operator is the standard.