NVIDIA's October 22 GPUDirect Storage tutorial identifies two independent routing decisions beneath a Python read. KvikIO provides Python and C++ abstractions above cuFile, but it can choose either cuFile or a POSIX compatibility path. If it chooses cuFile, a second decision remains: cuFile can use the accelerated nvidia-fs route or a CPU-staged compatibility route, depending on support and configuration.
Successfully calling KvikIO—or even establishing that it called cuFile—therefore does not demonstrate that the accelerated path moved the data. The interface can preserve correctness while the physical route changes.
NVIDIA recommends enabling CUFILE_STATS_LEVEL=3 and inspecting gds_stats. The accelerated route appears as nvfs; the compatibility route appears as posix. In its example, a 1 GiB read divided into 4 MiB tasks generated 256 nvfs operations. Both KvikIO and cuFile can decompose work, so application task counts and lower-level operation counts need interpretation.
The example used an H100, CUDA 13.3, cuFile 1.18 and local NVMe storage on ext4, with IOMMU and ACS disabled and nvidia-fs loaded. Those are the conditions of that demonstration, not a general instruction to disable platform controls or a performance result for other storage and topology combinations.
LEON's inference is that GPU-storage benchmark provenance should include the actual route, device, filesystem, topology, fallback behavior and decomposition. A throughput result alone cannot establish which architectural hypothesis was tested. Two machines can run code against the same library while moving data through materially different paths.
A compatibility fallback need not be treated as an error. It may be entirely appropriate for a research job whose first requirement is completion. But a system designed around a CPU budget or latency assumption needs to know when the fallback occurred. The observation should be available before the workload owner chooses whether to accept, flag or exclude that run. The tutorial supplies a practical route check without establishing a universal performance penalty or speedup.