# =====================================================================
# Experiment: Prefill-Decode Disaggregation Load Curve
# =====================================================================
#
# Measures how a disaggregated prefill-decode architecture handles
# increasing concurrency.  Each treatment scales concurrency and the
# number of prompts proportionally (10 prompts per concurrency slot)
# to keep per-worker load consistent while increasing overall system
# pressure.  This produces a load curve showing how latency and
# throughput change as the system saturates.
#
# Usage (full DoE — automatic standup/run/teardown per setup treatment):
#   llmdbenchmark --spec pd-disaggregation experiment \
#     --experiments workload/experiments/pd-disaggregation.yaml
#
# Usage (run-only — targets an already-running endpoint):
#   llmdbenchmark --spec pd-disaggregation run \
#     --endpoint-url http://<HOST>:<PORT> \
#     --model <MODEL_NAME> \
#     --namespace <NAMESPACE> \
#     --harness vllm-benchmark \
#     --workload random_concurrent.yaml \
#     --experiments workload/experiments/pd-disaggregation.yaml
#
# The experiment form auto-detects the endpoint from stood-up stacks
# and reads harness/profile defaults from the plan config.
# The run-only form requires explicit --endpoint-url, --model, and
# --namespace since there is no stood-up stack to detect from.
#
# =====================================================================

# --- DoE Metadata (informational — not consumed by the runtime) ------

experiment:
  name: pd-disaggregation
  description: >
    Proportional load scaling curve for prefill-decode disaggregated
    inference.  Concurrency and prompt count scale together (ratio
    1:10) to measure system behavior from minimal load through
    saturation.
  harness: vllm-benchmark
  profile: random_concurrent.yaml

design:
  type: proportional_scaling

  setup:
    factors:
      - name: deploy_method
        key: deploy_method
        levels: [modelservice, standalone]
        description: Deployment method
      - name: decode_replicas
        key: decode.replicas
        levels: [1, 2, 3, 4]
        description: Number of decode replicas
      - name: decode_tensor_parallelism
        key: decode.parallelism.tensor
        levels: [2, 4, 8]
        description: Tensor parallelism for decode pods
      - name: prefill_replicas
        key: prefill.replicas
        levels: [2, 4, 6, 8]
        description: Number of prefill replicas
      - name: prefill_tensor_parallelism
        key: prefill.parallelism.tensor
        levels: [1, 2, 4]
        description: Tensor parallelism for prefill pods

  run:
    factors:
      - name: max-concurrency
        key: max-concurrency
        levels: [1, 8, 32, 64, 128, 256]
        description: >
          Maximum number of concurrent in-flight requests.  Controls
          the degree of parallelism hitting the serving endpoint.

      - name: num-prompts
        key: num-prompts
        levels: [10, 80, 320, 640, 1280, 2560]
        description: >
          Total number of prompts to send.  Scaled proportionally with
          concurrency (10x) so each concurrency slot processes roughly
          the same number of requests.

    factor_relationship: >
      num-prompts = max-concurrency * 10.  This proportional scaling
      keeps the per-worker load constant across treatments, isolating
      the effect of concurrency on system throughput and tail latency.

    constants:
      - key: dataset-name
        value: random
        description: Random synthetic dataset
      - key: random-input-len
        value: 10000
        description: Input token length (held fixed at 10K)
      - key: random-output-len
        value: 1000
        description: Output token length (held fixed at 1K)

  total_setup_treatments: 9
  total_run_treatments: 6
  total_matrix: 54

  response_variables:
    - TTFT (Time to First Token)
    - TPOT (Time per Output Token)
    - ITL (Inter-Token Latency)
    - E2EL (End-to-End Latency)
    - Throughput (tokens/sec)
    - p99 / p99.9 latency percentiles

# --- Runtime: Setup treatments (consumed by experiment orchestrator) -------
#
# Fractional factorial: 9 selected topologies from the full factor space.

setup:
  treatments:
    - name: ms-d1xTP8-p8xTP1
      decode.replicas: 1
      decode.parallelism.tensor: 8
      prefill.replicas: 8
      prefill.parallelism.tensor: 1

    - name: ms-d1xTP8-p4xTP2
      decode.replicas: 1
      decode.parallelism.tensor: 8
      prefill.replicas: 4
      prefill.parallelism.tensor: 2

    - name: ms-d1xTP8-p2xTP4
      decode.replicas: 1
      decode.parallelism.tensor: 8
      prefill.replicas: 2
      prefill.parallelism.tensor: 4

    - name: ms-d1xTP4-p6xTP2
      decode.replicas: 1
      decode.parallelism.tensor: 4
      prefill.replicas: 6
      prefill.parallelism.tensor: 2

    - name: ms-d2xTP4-p4xTP2
      decode.replicas: 2
      decode.parallelism.tensor: 4
      prefill.replicas: 4
      prefill.parallelism.tensor: 2

    - name: ms-d3xTP4-p2xTP2
      decode.replicas: 3
      decode.parallelism.tensor: 4
      prefill.replicas: 2
      prefill.parallelism.tensor: 2

    - name: sa-d1xTP2
      standalone.enabled: true
      decode.replicas: 1
      decode.parallelism.tensor: 2

    - name: sa-d1xTP4
      standalone.enabled: true
      decode.replicas: 1
      decode.parallelism.tensor: 4

    - name: sa-d1xTP8
      standalone.enabled: true
      decode.replicas: 1
      decode.parallelism.tensor: 8

# --- Treatments (consumed by step_04 render_profiles) ----------------
#
# Proportional scaling: concurrency and num-prompts scale together
# at a fixed 1:10 ratio.  6 treatments from light to heavy load.
#
# Naming: conc{max-concurrency}

treatments:
  - name: conc1
    max-concurrency: 1
    num-prompts: 10

  - name: conc8
    max-concurrency: 8
    num-prompts: 80

  - name: conc32
    max-concurrency: 32
    num-prompts: 320

  - name: conc64
    max-concurrency: 64
    num-prompts: 640

  - name: conc128
    max-concurrency: 128
    num-prompts: 1280

  - name: conc256
    max-concurrency: 256
    num-prompts: 2560