Add "gpu_fakedev" documentation

Signed-off-by: Eero Tamminen <[email protected]>

Author: eero-t

cmd/gpu_fakedev/README.md

@@ -0,0 +1,47 @@
+# Fake (GPU) device file generator
+
+Table of Contents
+* [Introduction](#introduction)
+* [Configuration](#configuration)
+* [Potential improvements](#potential-improvements)
+* [Related tools](#related-tools)
+
+## Introduction
+
+This is a tool for generating (large number of) fake device files for
+k8s device scheduling scalability testing.  But it can also be used
+just to test (GPU) device plugin functionality without having
+corresponding device HW.
+
+Its "intel-gpu-fakedev" container is intended to be run as first init
+container in a device plugin pod, so that device plugin (and its NFD
+labeler) see the fake (sysfs + devfs) files generated by the tool,
+instead of real host sysfs and devfs content.
+
+## Configuration
+
+[Configs](configs/) subdirectory contains example JSON configuration
+file(s) for the generator. Currently there's only one example JSON
+file, but each new device variant adding feature(s) that have specific
+support in device plugin, could have their own fake device config.
+
+## Potential improvements
+
+If support for mixed device environment is needed, tool can be updated
+to use node / configuration file mapping.  Such mappings could be e.g.
+in configuration files themselves as node name include / exlude lists,
+and tool would use first configuration file matching the node it's
+running on.  For now, one would need to use different pod / config
+specs for different nodes to achieve that...
+
+Currently JSON config file options and the generated files are tied to
+what GPU plugin uses, but if needed, they could be changed to fake
+also sysfs + devfs device files used by other plugins.
+
+## Related tools
+
+[fakedev-exporter](#https://github.com/intel/fakedev-exporter) project
+can be used to schedule suitably configured fake workloads on the fake
+devices, and to provide provide fake activity metrics for them to
+Prometheus, that look like they were reported by real Prometheus
+metric exporters for real workloads running on real devices.