net.thevpc.nuts.runtime.standalone.platform

Class NLinuxGpuProbe

java.lang.Object

net.thevpc.nuts.runtime.standalone.platform.NLinuxGpuProbe

public class NLinuxGpuProbe
extends Object

Discovers GPU devices on linux, as NGpuDevice instances carrying the capabilities declared by NGpuDevice.

Detection is layered by cost. The kernel exposes vendor, pci identity, bound module, device count, model name and uuid as plain files, and those readings are identical on every distribution because they are kernel interfaces rather than a userspace layout. Only compute capability and NVIDIA memory amounts require running nvidia-smi.

The identity of a device cannot change while the machine is up, so it is probed once and cached. Memory amounts can and do change constantly, so they are re-read on every call and never cached : a caller asking how much video memory is free must not be handed a figure from process startup.

The filesystem root and the command runner are injectable, so that the whole reading can be exercised against fixture directories rather than against the machine the tests happen to run on.

Every probe is fail-safe, an unreadable or missing source degrades the corresponding property to unknown and never propagates an error : a machine with no GPU, a container without device passthrough and a card left to nouveau are all normal outcomes rather than failures.

Method Summary

Modifier and Type	Method and Description
List<NGpuDevice>	detect() Reads the devices under this probe's root.
static List<NGpuDevice>	gpus() Probes the machine.
static NLinuxGpuProbe	of(Path root, Function<String[],String> commandRunner) Builds a probe reading an arbitrary root through an arbitrary command runner, which is what makes the reading testable : a fixture directory standing in for /sys and /proc, and a stub standing in for the vendor tool.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Method Detail

of

public static NLinuxGpuProbe of(Path root,
                                Function<String[],String> commandRunner)

Builds a probe reading an arbitrary root through an arbitrary command runner, which is what makes the reading testable : a fixture directory standing in for /sys and /proc, and a stub standing in for the vendor tool.

The runner is handed the command and its arguments and returns the standard output, or null when the tool is missing, fails or times out, which is the contract exec(String...) honours.

Parameters:

root - directory standing in for the filesystem root

commandRunner - runner invoking a command, null result meaning no answer

Returns:

a probe reading that root

gpus

public static List<NGpuDevice> gpus()

Probes the machine.

Device identity is cached, memory amounts are read afresh on every call, so that the returned NRam describes the machine now rather than at first probe.

Returns:

detected devices sorted by pci address, empty when none is found

detect

public List<NGpuDevice> detect()

Reads the devices under this probe's root. Unlike gpus() it does not require the host to be linux, the root being whatever it was built with.

Returns:

detected devices sorted by pci address, empty when none is found