Define a useful start
A start measurement needs an observable finish. For an inference service, opening a port may precede model loading, initialization and the first valid response. Choose the moment at which the application can do its intended job, then divide the path into boot, service launch, artifact loading and first work. Retain the timestamps in a common record. This makes the result actionable: a slow download, a long model initialization and a slow storage read call for different changes, even if they all appear as a long wait to the person using the device.
Keep cache state visible
The second launch can reuse data or artifacts that were unavailable to the first. Describe whether the measurement follows a reboot, a process restart or a prior completed request. These are useful scenarios in their own right, but mixing them produces a number that is hard to interpret. Also note whether generated artifacts already exist and whether the model is local or must be retrieved. Those conditions can change startup without changing the underlying storage hardware at all. fio’s documentation distinguishes buffered and non-buffered I/O and provides controls for access patterns. A synthetic storage test should therefore state its own conditions; its result does not automatically describe the application’s mixture of reads, computation and initialization.
Use storage tests to answer a narrower question
If the startup trace points toward reading artifacts, inspect the size and shape of those reads before choosing a diagnostic workload. Use an explicitly identified scratch file for any synthetic test and preserve the job definition. Then repeat the application start after one storage change, holding the software and model fixed. Report startup improvement only when that end-to-end observation supports it. Storage capacity, interface speed and a sequential benchmark are different facts. The useful upgrade is the one that resolves the application’s identified delay within the complete system’s power, space and recovery requirements.
Sources & publication notes
Retrospective period: July 2025. Written and published November 2, 2026. Product announcements establish release timing; support pages are linked as current references.