Skip to content

planner-guide.md

Latest commit

 

History

History
242 lines (181 loc) · 14.9 KB

planner-guide.md

File metadata and controls

242 lines (181 loc) · 14.9 KB
title Planner Guide

简体中文

The Dynamo Planner is an autoscaling controller that adjusts prefill and decode engine replica counts at runtime to meet latency SLAs. It reads traffic signals (Prometheus metrics or load predictor output) and engine performance models to decide when to scale up or down.

For a quick overview, see the Planner overview. For architecture internals, see Planner Design.

Scaling Modes

The planner supports four optimization targets that determine how scaling decisions are made:

  • throughput (default): Uses static thresholds on queue depth and KV cache utilization. No SLA targets or profiling needed. Works out of the box.
  • latency: Same approach as throughput but with more aggressive thresholds — scales up earlier and tolerates less queuing. Ideal for latency-sensitive workloads.
  • load: Uses user-defined prefill queue token thresholds and decode KV utilization thresholds for reactive load-based scaling.
  • sla: Uses the Rust engine performance shim with native AIC estimates when available, plus online FPM tuning or FPM regression fallback, to target specific TTFT/ITL values. Supports both throughput-based (predictive) and load-based (reactive) scaling modes. For advanced users who need precise SLA control.

When to use which:

  • Start with throughput (the default) — it works immediately with no configuration.
  • Switch to latency if your workload has strict latency requirements and you prefer to over-provision rather than queue.
  • Use load when you want direct control through prefill queue and decode KV utilization thresholds.
  • Use sla when you want to target specific TTFT/ITL values with native AIC estimates, optional bootstrap profiling data, or live FPM warmup.

PlannerConfig Reference

The planner is configured via a PlannerConfig JSON/YAML object. When using the profiler, this is placed under the features.planner section of the DGDR spec:

features:
  planner:
    mode: disagg
    backend: vllm
    # optimization_target defaults to "throughput" — works out of the box

For SLA-based scaling:

features:
  planner:
    optimization_target: sla
    enable_throughput_scaling: true
    enable_load_scaling: false
    pre_deployment_sweeping_mode: rapid
    mode: disagg
    backend: vllm

To evaluate Planner behavior without changing replica counts, turn on advisory mode:

features:
  planner:
    advisory: true

Advisory mode is suggestion-only. The Planner computes recommended replica counts, logs them, exports them as diagnostics, and shows them in HTML reports. The recommendations are not applied as scaling decisions: the Planner does not execute scaling actions or change the deployment.

Optimization Target

Field Type Default Description
optimization_target string throughput throughput: scale based on queue/utilization thresholds. latency: aggressive low-latency thresholds. load: user-defined prefill queue and decode KV utilization thresholds. sla: Rust engine perf model scaling with ttft_ms/itl_ms targets.

When optimization_target is throughput, latency, or load, load-based scaling is automatically enabled and throughput-based scaling is disabled. The ttft_ms/itl_ms fields are ignored.

Scaling Mode Fields (SLA mode)

Field Type Default Description
enable_throughput_scaling bool true Enable throughput-based scaling. Only used when optimization_target: sla.
enable_load_scaling bool false Enable load-based scaling. Only used when optimization_target: sla.

At least one scaling mode must be enabled when using optimization_target: sla.

Pre-Deployment Sweeping

Field Type Default Description
pre_deployment_sweeping_mode string rapid How to generate optional bootstrap performance data: rapid (AIC simulation, ~30s), thorough (real GPUs, 2-4h), or none (skip).

SLA mode uses the Rust engine performance shim. If aic_perf_model is present, the planner initializes the shim with native AIC model identity and engine limits. Unsupported native AIC configs automatically fall back to observed-FPM regression in the shim. If aic_perf_model is absent, the shim starts as an FPM regression model and becomes ready after enough self-benchmark or live FPM observations.

At startup, the planner always tries to fetch self-benchmark results from the get_perf_metrics Dynamo endpoint. If unavailable, it falls back to rapid-mode AIC interpolation data or profiler-generated data (npz or JSON) at profile_results_dir when configured. These sources are converted to ForwardPassMetrics and used to tune or bootstrap the perf model. With pre_deployment_sweeping_mode: none, the planner can still start; throughput decisions report model_not_ready until native AIC is available or enough live FPMs have warmed the regression fallback.

Manual native AIC perf-model config:

features:
  planner:
    optimization_target: sla
    aic_perf_model:
      hf_id: nvidia/Llama-3.1-8B-Instruct-FP8
      system: h200_sxm
      backend: vllm
      prefill_pick: {tp: 1, pp: 1, dp: 1, moe_tp: 1, moe_ep: 1}
      decode_pick: {tp: 1, pp: 1, dp: 1, moe_tp: 1, moe_ep: 1}

Throughput-Based Scaling Settings

Field Type Default Description
throughput_adjustment_interval_seconds int 180 Seconds between throughput-based scaling decisions.
throughput_metrics_source string frontend Prometheus traffic source for throughput scaling: frontend reads dynamo_frontend_* metrics from the public Frontend; router reads dynamo_component_router_* metrics from a LocalRouter. Use router for pool-local Planner in GlobalPlanner deployments.
min_endpoint int 1 Minimum number of engine endpoints to maintain.
max_gpu_budget int 8 Maximum total GPUs the planner may allocate.
ttft_ms float 500.0 TTFT SLA target (ms) for scaling decisions.
itl_ms float 50.0 ITL SLA target (ms) for scaling decisions.

Load-Based Scaling Settings

Field Type Default Description
load_adjustment_interval_seconds int 5 Seconds between FPM tuning updates and load-based scaling decisions. Even when only throughput scaling is enabled, live FPM observations are fed into the perf model at this interval. Must be shorter than throughput_adjustment_interval_seconds.
max_num_fpm_samples int 64 Maximum retained FPM observations for online tuning or regression.
fpm_sample_bucket_size int 16 Number of buckets for observation retirement (must be a perfect square).
load_scaling_down_sensitivity int 80 Scale-down sensitivity 0–100 (0=never, 100=aggressive).
load_min_observations int 5 Minimum observations before making scaling decisions.

General Settings

Field Type Default Description
mode string disagg Planner mode: disagg, prefill, decode, or agg.
backend string vllm Backend: vllm, sglang, trtllm, or mocker.
environment string kubernetes Runtime environment: kubernetes, virtual, or global-planner.
namespace string env DYN_NAMESPACE Kubernetes namespace for the deployment.
advisory bool false Suggestion-only mode. Compute, log, export, and report recommended replica counts without executing scaling actions or changing the deployment.

Traffic Prediction Settings

Field Type Default Description
load_predictor string arima Prediction method for request count, ISL, and OSL: constant, arima, kalman, or prophet. Runtime metadata such as KV hit rate and speculative decode accept length uses the latest valid observation instead.
load_predictor_log1p bool false Apply log1p transform to predicted request count, ISL, and OSL data.
prophet_window_size int 50 Window size (seconds) for Prophet predictor.
load_predictor_warmup_trace string null Path to a warmup trace file for bootstrapping predictions.

KV hit rate and speculative decode accept length are runtime engine/router signals, not traffic shape. The planner stores the latest valid observation for each signal and reuses it until a newer valid value arrives. On cold start, missing KV hit rate means no prefix-cache discount, and missing accept length means 1.0.

Kalman Filter Settings

Field Type Default Description
kalman_q_level float 1.0 Process noise for level component.
kalman_q_trend float 0.1 Process noise for trend component.
kalman_r float 10.0 Measurement noise.
kalman_min_points int 5 Minimum data points before Kalman predictions activate.

Diagnostics Reports

Field Type Default Description
report_interval_hours float or null 24.0 Generate an HTML diagnostics report every N hours (simulated time). Set to null to disable periodic report generation.
report_output_dir string ./planner_reports Directory for HTML diagnostics reports.
live_dashboard_port int 8080 Port for the live diagnostics dashboard HTTP server. Set to 0 to disable. When enabled, visit http://host:port/ to view a real-time Plotly report of accumulated snapshots.

The same diagnostic signals surfaced in these reports are also exported as Prometheus metrics under the dynamo_planner_* prefix—for example estimated TTFT/ITL (dynamo_planner_estimated_ttft_ms, dynamo_planner_estimated_itl_ms), recommended replica counts (dynamo_planner_predicted_num_prefill_replicas, dynamo_planner_predicted_num_decode_replicas), per-engine capacity and FPM queue depths, and load/throughput scaling decision enums.

The Replica Counts plot overlays actual prefill/decode replicas with discrete recommendation markers for the Planner's recommended prefill/decode replicas. When advisory: true, these recommended counts are suggestions only; the Planner records what it would do without applying the change.

Scheduling / plugin pipeline

The planner runs through the builtin plugin pipeline by default. The base pipeline cadence lives under the scheduling sub-tree of PlannerConfig; plugin registration, transport, and auth settings live under plugin_registration.

Field Type Default Description
scheduling.scale_interval_seconds float gcd of enabled builtin intervals Base pipeline cadence. The pipeline wakes once per interval; each plugin's execution_interval_seconds decides whether that plugin fires on the tick. By default, the cadence is the gcd of load_adjustment_interval_seconds and, when throughput scaling is enabled, throughput_adjustment_interval_seconds, preserving existing config fire times.
scheduling.tick_max_duration_seconds float 30.0 Outer deadline wrapping the full plugin pipeline. Exceeding it aborts the tick; the next tick runs from a clean state.
plugin_registration.transport.request_timeout_seconds float 5.0 Per-plugin RPC timeout. Plugins exceeding this raise PluginTimeoutError; the stage continues with the remaining plugins.

Existing planner fields still drive the builtin plugins:

  • load_adjustment_interval_seconds schedules builtin_load_propose, which reads FPM and worker-count observations and applies the current load-based algorithm.
  • throughput_adjustment_interval_seconds schedules builtin_load_predict and builtin_throughput_propose. The throughput proposer requires the prediction from the same tick, so it only fires when the predict plugin fires.
  • When both builtins propose targets in the same tick, load-based scaling runs after throughput-based scaling and preserves the existing behavior: throughput updates the lower-bound replicas, then load-based scaling can adjust above that floor and apply the global GPU budget clamp.
  • After the plugin pipeline finishes, the planner applies the same final min_endpoint and GPU-budget safety checks to builtin and external-plugin targets before scaling the deployment.

DGDR example

apiVersion: nvidia.com/v1beta1
kind: DynamoGraphDeploymentRequest
metadata:
  name: my-deployment
spec:
  model: Qwen/Qwen3-0.6B
  features:
    planner:
      optimization_target: sla
      enable_load_scaling: true
      ttft: 200.0
      itl: 10.0
      pre_deployment_sweeping_mode: rapid
      scheduling:
        tick_max_duration_seconds: 30.0
      plugin_registration:
        transport:
          request_timeout_seconds: 5.0

Integration with Profiler

When the profiler runs with planner enabled, it:

  1. Selects the best prefill and decode engine configurations
  2. Generates engine performance data (prefill TTFT vs ISL, decode ITL vs KV-cache utilization)
  3. Saves the PlannerConfig and performance data into separate Kubernetes ConfigMaps
  4. Adds the planner service to the generated DGD, configured to read from those ConfigMaps

The planner receives its config via --config /path/to/planner_config.json which is mounted from the planner-config-XXXX ConfigMap. When thorough bootstrap data is generated, profiling data is mounted from the planner-profile-data-XXXX ConfigMap.

See the Profiler Guide for the full profiling workflow and how to configure pre-deployment sweeping.

Hierarchical Deployments

If you want one public endpoint for a model but multiple private DGDs optimized for different request classes, use a hierarchical deployment:

  • one control DGD with Frontend, GlobalRouter, and GlobalPlanner
  • one or more prefill pool DGDs
  • one or more decode pool DGDs

In the current workflow, run profiling independently for each intended pool, then compose the final control DGD plus pool DGDs manually. See the Global Planner Guide.

See Also