fix: Dynamic GPU partitioning lacks single-GPU-level granularity. (#1…

[Goend]

…054)

What type of PR is this?

What this PR does / why we need it:
fix: Dynamic GPU partitioning lacks single-GPU-level granularity. (#1…

Which issue(s) this PR fixes:
Fixes #1054

Special notes for your reviewer:

Does this PR introduce a user-facing change?:

[archlitchi]

/assign @ouyangluwei163

[ouyangluwei163]

migConfigs["current"][idx].MigDevices is Map, Map types are reference types. https://github.com/Project-HAMi/HAMi/blob/master/pkg/device-plugin/nvidiadevice/nvinternal/plugin/util.go#L255C8-L255C73

[Goend]

Is the intention here to perform a deep copy of the map type instead of using a reference?

[ouyangluwei163]

yes, if multiple cards in the machine are one flavor.

version: v1
mig-configs:
  current:
  - devices: [0, 1, 2]
    mig-enabled: true
    mig-devices:
      1g.5gb: 1
      2g.10gb: 3

and if not deep copy, when device 0 is modified, devices 1/2 will also be modified

[Goend]

done.the deepCopyMigConfig function has been used to perform a deep copy

[ouyangluwei163]

great, I will test it later

[ouyangluwei163]

Here you can simplify, 692-715

		for i := 0; i < deviceCount; i++ {
			var found bool
			for idx, config := range migConfigs["current"] {
				if len(config.Devices) == 0 || containsDevice(i, config.Devices) {
					newConfig := migConfigs["current"][idx]
					newConfig.MigDevices = make(map[string]int32)
					for k, v := range migConfigs["current"][idx].MigDevices {
						newConfig.MigDevices[k] = v
					}
					newConfig.Devices = []int32{int32(i)}
					result = append(result, newConfig)
					found = true
					break
				}
			}
			if !found {
				return nil, fmt.Errorf("device %d does not match any MIG configuration", i)
			}
		}

[Goend]

done,use additional auxiliary structures

deviceToConfig := make(map[int32]*nvidia.MigConfigSpec)

to reduce the number of for loop layers

[Goend]

The error here does not need to be handled. If the corresponding device configuration is not found, simply continue and issue a warning. There is no need for
return nil, fmt.Errorf("device %d does not match any MIG configuration", i)
instead, just use continue. and return other config
such as return config,fmt.Errorf("device %d does not match any MIG configuration", i)

[Goend]

@ouyangluwei163 What do you think?

[ouyangluwei163]

I think it's OK

[archlitchi]

@ouyangluwei163 ready to integrate?

[Goend]

when signal gpu in node,the generate config is ok

I0523 07:36:58.771205   58171 util.go:204] Applying data= version: v1
mig-configs:
    current:
        - devices: [0]
          mig-enabled: true
          mig-devices:
            1g.5gb: 7

F0523 07:36:59.384536   58171 util.go:212] nvidia-mig-parted failed with exit status 1
[root@node-37 ~]# kubectl  -n hami  logs hami-device-plugin-42wt2 --previous|grep test
Defaulted container "device-plugin" out of: device-plugin, vgpu-monitor
I0523 07:36:52.418380   58171 server.go:241] test

[ouyangluwei163]

/lgtm

[archlitchi]

/lgtm

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Flag	Coverage Δ
unittests	61.14% <ø> (+0.07%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

see 1 file with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.