NVIDIA explains how to verify the path into GPU memory

NVIDIA’s new tutorial covers storage reads into graphics-processor memory. GPUDirect Storage can avoid staging data in CPU memory on supported systems. However, calling its cuFile interface does not prove that this direct route was used. The walkthrough requires Linux, an NVIDIA GPU, installed GPUDirect Storage, supported storage and filesystem configuration, and permission to access files and diagnostic tools. Its cuFile statistics distinguish accelerated nvidia-fs transfers from CPU-staged compatibility transfers.

The runtime documentation explains the first choice made by KvikIO. Its AUTO mode tries cuFile and falls back to conventional POSIX file access when requirements are unmet. ON selects that compatibility route directly. OFF enforces cuFile, but operations can fail on an unsuitable system. Crucially, cuFile has its own compatibility mode: selecting the library is a separate decision from the transfer route it ultimately uses.

The statistics guide adds a way to inspect the application’s work. Creating a SummaryMonitor starts collecting statistics; its get method reads accumulated totals. By default, an operation means one application call, even if KvikIO splits it into smaller transfers. Physical observation instead counts those transfers. The guide also distinguishes busy bandwidth from throughput over the entire elapsed time: overlapping operations count once as busy time, while gaps between calls remain idle.