Running AI Workloads on Rack-Scale Supercomputers: From Hardware to Topology-Aware Scheduling

NVIDIA is addressing the operational complexity of its new Blackwell-based rack-scale systems with a software stack designed to make hardware topology comprehensible to workload schedulers. The initiative, centered around tools like NVIDIA Mission Control and Topograph, aims to ensure that large-scale AI workloads running on GB200 and GB300 NVL72 systems achieve expected performance. The core issue is that schedulers such as Slurm and Kubernetes are traditionally unaware of the sophisticated, high-bandwidth NVLink fabric connecting GPUs, which can lead to inefficient job placement and performance bottlenecks. This software provides the necessary translation layer to align application scheduling with the physical layout of the supercomputer.

The technical solution hinges on exposing the system's NVLink topology to schedulers through two key identifiers: a 'cluster UUID' for the entire NVLink domain (the rack) and a 'clique ID' for specific NVLink partitions within it. Mission Control centralizes this information, allowing schedulers to make informed placement decisions. For Slurm, this integration is achieved via the 'topology/block' plugin, which treats high-bandwidth NVLink partitions as distinct blocks for job allocation. In the Kubernetes ecosystem, the NVIDIA DRA (Dynamic Resource Allocation) driver introduces a concept called 'ComputeDomains,' which groups nodes sharing an NVLink domain, ensuring that distributed pods are co-located for high-performance communication. NVIDIA's Run:ai platform further abstracts this complexity, automating the creation of ComputeDomains and handling topology-aware placement for users.

This development signals a significant move towards providing integrated, full-stack solutions for enterprise AI infrastructure. By creating a robust software control plane, NVIDIA is making its most powerful hardware more manageable and accessible for AI architects and HPC platform operators. The impact is a reduction in operational overhead and a greater ability to deliver predictable performance and resource isolation in multi-tenant AI environments. This focus on software-defined infrastructure management is critical for turning collections of powerful hardware into functional and efficient 'AI factories' capable of handling the next generation of complex models.

NVIDIA is shifting focus from purely marketing hardware speeds to providing the critical software control plane for managing it, signaling that operationalizing rack-scale AI infrastructure is now as important as the underlying silicon itself.