NVIDIA’s GDS Diagnostic Tool Turns Storage Tuning Into a Guided Mission

GPUDirect Storage can transform the path between NVIDIA GPUs and storage targets, but getting that path configured correctly takes more than installing software. NVIDIA’s open-source gds-diag gives administrators a direct way to inspect their environments, find blockers, and plan improvements before performance problems spread across a workload.

The tool arrives with a second option for teams that want a more conversational experience: an agent skill that understands natural-language requests, coordinates multiple diagnostic steps, and uses model-guided reasoning to interpret results. Whether you prefer a command line or an AI coding assistant, the goal is the same—make GDS configuration easier to check, explain, and improve.

A Clearer Path Through GPUDirect Storage

GDS accelerates GPU storage access in on-premises and cloud environments by creating a data path between NVIDIA GPUs and storage targets. That path is designed to increase bandwidth, reduce latency, and lower CPU and memory utilization across different platforms, application frameworks, and workload patterns.

Yet a fast data path depends on the surrounding platform. GPU support, CUDA Toolkit installation, configuration files, environment variables, storage targets, and container settings can all influence whether GDS works as intended. The GPUDirect Storage team developed gds-diag to simplify that optimization and configuration work.

gds-diag is a single Python script with multiple subcommands, each aimed at a different diagnostic task. Its checks cover the full journey from pre-installation planning to post-installation verification, with deeper options for containers and specific mount points.

Every finding arrives with a status badge such as FAIL, WARN, INFO, or PASS. The result does not stop at a label: each finding includes an explanation and a mitigation plan that administrators can put into action.

Installation and the First Diagnostic Run

Using gds-diag requires a Linux system with Python 3.8 or newer. The utility itself comes through the open-source NVIDIA Magnum IO GitHub repository, with no other dependencies to install. GPUDirect Storage has separate platform requirements: an NVIDIA datacenter-class GPU and NVIDIA CUDA Toolkit 12.2 or newer, but those requirements are not dependencies of gds-diag itself.

Start by cloning the NVIDIA Magnum IO repository on the machine where you plan to run the utility. Then move into the gds-diag directory, run gds-diag.py –help, and follow the command-line usage guidance.

The help output lists the available subcommands. For details about one specific command, run ./gds-diag.py <subcommand> –help. If you are not sure where to begin, the broadest starting point is ./gds-diag.py all /path/to/check, which runs a complete sequence of diagnostics for that path.

That starting point works across bare-metal, virtual machine, and container environments, giving teams one entry point before they move into more focused checks.

Four Diagnostic Paths, One Operating Picture

The available subcommands fall into four groups, each answering a different question about the GDS environment:

• End-to-end checks: The all command selects suitable diagnostics for the environment and runs them in sequence.
• Preparing for installation: The pre-install command checks GPU support and CUDA Toolkit installation, while support-matrix shows supported storage targets for deployment planning.
• Validating an installation: The post-install command checks whether GDS is installed correctly and identifies blockers. The config-audit command reviews configuration files and environment variables that influence GDS behavior.
• Deep diagnostics: The container-check command assesses GDS support in container environments, including containerd, Docker, Enroot, and Kubernetes.

This structure lets an administrator match the command to the stage of a deployment. A team preparing a new system can check support before installation, while a team troubleshooting an existing system can review configuration, installation status, or container support without running every available test.

An AI Agent for Multi-Step Troubleshooting

The optional agent skill adds a natural-language interface to the same diagnostic work. Open the gds-diag directory in a preferred AI coding assistant, then follow the Agent skill usage guidance. The skill interprets intent, orchestrates multi-step workflows, and proposes remediation based on the results.

The AI agent mode is recommended when it is available in your environment. gds-diag has been validated with Claude and Codex, though it may work with other tools as well. Both the command-line utility and the agent skill are distributed through the NVIDIA Magnum IO GitHub repository under open-source licenses.

That combination changes the starting question from “Which command should I run?” to “What do I need to check?” The agent can connect related findings across the diagnostic workflow, while the status badges and mitigation plans keep the final recommendations visible to administrators.

GPUDirect Storage sits at the intersection of GPU hardware, CUDA software, storage targets, configuration, and deployment environments. With gds-diag, those layers become a sequence of checks instead of a maze of isolated settings. As GDS expands across on-premises and cloud platforms, a repeatable way to verify the path could become as important as the path itself.

Stay connected via Google News

Follow us for the latest AI News