Not showing warning when using a taken GPU

[ruspaul013]

Nomad version

nomad server: Nomad v1.6.1 + patch with #18141 nomad client: Nomad v1.6.1 + patch with #18141

Operating system and Environment details

Plugin "nomad-driver-podman" v0.5.1 Plugin "nomad-device-nvidia" v1.0.0

Issue

I created a patch with the solution provides in #18141 to test on our cluster. While testing I discovered that if a GPU or multiple GPU are already used in jobs, nomad will not give a warning about this and will place the job without using those GPUs.

Reproduction steps

1. Create a job file where you use multiple GPUs and set constraint based on their UUID.
2. Create a new job file where you use one or more GPUs that is already used in the first job.

Expected Result

Throw an warning like WARNING: Failed to place all allocations.

Actual Result

Place the job on the client.

Job file (if appropriate)

Job file 1:

job "test-2070-2" {
  datacenters = ["dc1"]
  group "test-2070-2" {

    restart {
        attempts=0
    }
    count=1
    task "test-2070-2" {
        driver = "podman"
        config {
            image = "image_with_gpu"
        }
        resources {
            cpu = 2650
            memory = 8192
            device "nvidia/gpu" {
                count = 2

                constraint {
                    attribute = "${device.model}"
                    value     = "NVIDIA GeForce RTX 2070 SUPER"
                }

                constraint {
                    attribute = "${device.ids}"
                    operator  = "set_contains"
                    value     = "GPU-9b5df054-6f08-f35c-9c4c-5709b19efea5,GPU-1846fc5f-8c71-bfab-00e1-9c190dd88ed7"
                }

            }
        }
    }
  }
}

Job file 2:

job "test-2070-2" {
  datacenters = ["dc1"]
  group "test-2070-2" {

    restart {
        attempts=0
    }
    count=1
    task "test-2070-2" {
        driver = "podman"
        config {
            image = "image_with_gpu"
        }
        resources {
            cpu = 2650
            memory = 8192
            device "nvidia/gpu" {
                count = 1

                constraint {
                    attribute = "${device.model}"
                    value     = "NVIDIA GeForce RTX 2070 SUPER"
                }

                constraint {
                    attribute = "${device.ids}"
                    operator  = "set_contains"
                    value     = "GPU-9b5df054-6f08-f35c-9c4c-5709b19efea5"
                }

            }
        }
    }
  }
}

[tgross]

@ruspaul013 I'm surprised that job spec doesn't simply return a validation error, but the constraint block belongs under job, group, or task, not under resources.device. Can you verify this is not working after that's corrected?

[ruspaul013]

Hello @tgross, thank you for the suggestion. I tried it, but now I get an error even if I do not use constraint block for ids. I can not place a simple job using a GPU.

resources {
      cpu = 3200*4
      memory = 8192
      device "nvidia/gpu" {
          count = 1

      }
  }

  constraint {
      attribute = "${device.model}"
      value     = "NVIDIA GeForce RTX 4090"
  }

and I get this warning

Scheduler dry-run:
- WARNING: Failed to place all allocations.
  Task Group "paulr_gpu_test" (failed to place 1 allocation):
    * Constraint "${device.model} = NVIDIA GeForce RTX 4090": 2 nodes excluded by filter

but the constraint block belongs under job, group, or task, not under resources.device

I know that constraint block does not belong there, but the example from device block says otherwise.

[lgfa29]

Yeah, that's something I often forget but you can set constraint and affinity at the device level 😅 https://developer.hashicorp.com/nomad/docs/job-specification/device#constraint

But setting the constraint only filters out nodes from scheduling. Taking a look at the code I think the problem is that AllocsFit doesn't take current device usage into account.

Devices are not considered comparable resources, so they're not filtered here: https://github.com/hashicorp/nomad/blob/e3c8700ded891702b7f94636033109efd8f71c3a/nomad/structs/funcs.go#L172-L178

And this part of the code only looks for device oversubscription among the allocs being scheduled, so it ignores the ones already running in the node: https://github.com/hashicorp/nomad/blob/e3c8700ded891702b7f94636033109efd8f71c3a/nomad/structs/funcs.go#L201-L207