Recipes Overview¶

Recipes are YAML configuration files that define how services and clients are deployed on the HPC cluster.

What Are Recipes?¶

Recipes provide a declarative way to specify:

• Container configuration: Image source and path
• Resource allocation: CPU, memory, GPU, time limits
• Environment variables: Service-specific settings
• Network ports: Exposed endpoints
• Monitoring: cAdvisor integration

Recipe Types¶

Directory Structure¶

recipes/
├── services/
│   ├── ollama.yaml
│   ├── ollama_with_cadvisor.yaml
│   ├── redis.yaml
│   ├── redis_with_cadvisor.yaml
│   ├── chroma.yaml
│   ├── chroma_with_cadvisor.yaml
│   ├── mysql.yaml
│   ├── mysql_with_cadvisor.yaml
│   ├── prometheus_with_cadvisor.yaml
│   └── grafana.yaml
└── clients/
    ├── ollama_benchmark.yaml
    ├── ollama_parametric.yaml
    ├── redis_benchmark.yaml
    ├── redis_parametric.yaml
    ├── chroma_benchmark.yaml
    ├── chroma_parametric.yaml
    └── mysql_benchmark.yaml

Basic Recipe Structure¶

Service Recipe¶

service:
  name: service_name
  description: "Human-readable description"

  # Container image
  container:
    docker_source: docker://image:tag
    image_path: $HOME/containers/image.sif

  # SLURM resources
  resources:
    nodes: 1
    ntasks: 1
    cpus_per_task: 4
    mem: "16G"
    time: "02:00:00"
    partition: gpu
    qos: default
    gres: "gpu:1"

  # Environment variables
  environment:
    VAR_NAME: "value"

  # Network
  ports:
    - 8080

  # Monitoring (optional)
  enable_cadvisor: true
  cadvisor_port: 8080

Client Recipe¶

client:
  name: client_name
  type: benchmark_type
  description: "Benchmark description"

  # Benchmark parameters
  parameters:
    param1: value1
    param2: value2
    output_file: "$HOME/results/output.json"

  # SLURM resources
  resources:
    cpus_per_task: 2
    mem: "4G"
    time: "00:30:00"
    partition: cpu

Using Recipes¶

Start a Service¶

python main.py --recipe recipes/services/ollama.yaml

Run a Client¶

python main.py --recipe recipes/clients/ollama_benchmark.yaml --target-service ollama_xxx

With Verbose Output¶

python main.py --verbose --recipe recipes/services/redis.yaml

Recipe Processing Flow¶

sequenceDiagram
    participant User
    participant CLI as main.py
    participant Loader as RecipeLoader
    participant Factory as JobFactory
    participant Service as ServiceClass
    participant Generator as ScriptGenerator
    participant SSH

    User->>CLI: --recipe service.yaml
    CLI->>Loader: load(path)
    Loader-->>CLI: recipe dict
    CLI->>Factory: create_service(recipe)
    Factory->>Service: __init__(recipe)
    Service-->>Factory: instance
    Factory-->>CLI: service
    CLI->>Generator: generate_script(service)
    Generator-->>CLI: slurm_script
    CLI->>SSH: submit_job(script)
    SSH-->>CLI: job_id
    CLI-->>User: Service started

Validation¶

Recipes are validated for:

• Required fields (name, container)
• Valid resource values
• Correct YAML syntax
• Environment variable format

Next: Service Recipes | Client Recipes