FLARE implementation

[prometherion]

No description provided.

[rmedina97]

I tried testing this but noticed that the Flavor doesn’t reflect some values from the FLARE annotations (GPU count and GPU memory stay at 0). Are the annotations the same as in this repository https://github.com/clastix/flare? For testing, should I follow the quickstart from that repo, or are there specific steps/examples for this PR to make validation easier?

[prometherion]

Are the annotations the same as in this repository https://github.com/clastix/flare?

Yes, I implemented the required code following that documentation: just sharing the annotations and the resulting flavor.

// kubectl get nodes fluidos-provider-2-worker -ojsonpath='{.metadata.annotations}'
{
  "cost.fluidos.eu/currency": "EUR",
  "cost.fluidos.eu/hourly-rate": "2.1",
  "gpu.fluidos.eu/architecture": "ampere",
  "gpu.fluidos.eu/clock-speed": "1.80G",
  "gpu.fluidos.eu/compute-capability": "8.6",
  "gpu.fluidos.eu/cores": "10752",
  "gpu.fluidos.eu/count": "8",
  "gpu.fluidos.eu/dedicated": "true",
  "gpu.fluidos.eu/fp32-tflops": "38.7",
  "gpu.fluidos.eu/interconnect": "nvlink",
  "gpu.fluidos.eu/interconnect-bandwidth-gbps": "600",
  "gpu.fluidos.eu/interruptible": "false",
  "gpu.fluidos.eu/memory-per-gpu": "48Gi",
  "gpu.fluidos.eu/model": "nvidia-a6000",
  "gpu.fluidos.eu/multi-gpu-efficiency": "0.85",
  "gpu.fluidos.eu/sharing-capable": "false",
  "gpu.fluidos.eu/sharing-strategy": "none",
  "gpu.fluidos.eu/tier": "standard",
  "gpu.fluidos.eu/topology": "ring",
  "gpu.fluidos.eu/vendor": "nvidia",
  "kubeadm.alpha.kubernetes.io/cri-socket": "unix:///run/containerd/containerd.sock",
  "location.fluidos.eu/region": "us-east-1",
  "location.fluidos.eu/zone": "zone-a",
  "network.fluidos.eu/bandwidth-gbps": "25",
  "network.fluidos.eu/latency-ms": "5",
  "network.fluidos.eu/tier": "standard",
  "node.alpha.kubernetes.io/ttl": "0",
  "provider.fluidos.eu/name": "cloud-provider-2",
  "provider.fluidos.eu/preemptible": "false",
  "volumes.kubernetes.io/controller-managed-attach-detach": "true",
  "workload.fluidos.eu/graphics-score": "0.95",
  "workload.fluidos.eu/hpc-score": "0.80",
  "workload.fluidos.eu/inference-score": "0.90",
  "workload.fluidos.eu/training-score": "0.85"
}

// kubectl get flavors.nodecore.fluidos.eu fluidos.eu-k8slice-89ad -ojsonpath='{.spec.flavorType.typeData.characteristics.gpu}'|jq
{
  "architecture": "ampere",
  "clock_speed": "1800M",
  "compute_capability": "8.6",
  "cores": "86016",
  "count": 8,
  "dedicated": true,
  "fp32_tflops": 38.7,
  "graphics_score": 0.95,
  "hourly_rate": 2.1,
  "hpc_score": 0.8,
  "inference_score": 0.9,
  "interconnect": "nvlink",
  "interconnect_bandwidth": "600",
  "memory": "384Gi",
  "model": "nvidia-a6000",
  "multi_gpu_efficiency": "0.85",
  "network_bandwidth": "25",
  "network_latency_ms": 5,
  "network_tier": "standard",
  "provider": "cloud-provider-2",
  "region": "zone-a",
  "sharing_strategy": "none",
  "tier": "standard",
  "topology": "ring",
  "training_score": 0.85,
  "vendor": "nvidia",
  "zone": "us-east-1"
}

For testing, should I follow the quickstart from that repo, or are there specific steps/examples for this PR to make validation easier?

Yes, we're going to release the code before the end of the week: testing will be easier.