Skip to content

About

No description, website, or topics provided.

Resources

Readme
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository
BranchesTags

Latest commit

 

History

7 Commits

Folders and files

Repository files navigation

OS-aware EDA-Lite

Lightweight MobileViT simulation for ROS2 OS-level edge AI performance analysis.

本專案是作業系統期末專題實作。目標不是完整重現 EDA compiler，而是建立一個可解釋、可重現的 EDA-inspired 輕量化 simulator，用來觀察 SRAM 限制、layer fusion 策略、ROS2 runtime pressure 與 DDS/掉幀壓力如何影響 edge AI 推論的估計成本。

目前主線是 MobileViT simplified profile simulation + ROS2 runtime pressure analysis。LLM 內容作為背景動機；本 repo 目前不執行真實 LLM inference。YOLO 相關路徑保留為早期 legacy work，不作為主流程。

Project Scope

目前已完成：

  • 讀取簡化模型 profile JSON。
  • 支援 MobileViT-XXS 與 MobileViT-S 的 coarse block simulation profiles。
  • 支援單一模型與多模型 profile sweep。
  • 支援三種排程策略：
    • layerwise：每層獨立執行。
    • greedy：由左到右盡量融合可行 layer。
    • dp：使用 dynamic programming 找最低估計成本的融合切分。
  • 支援 SRAM sweep：256KB、512KB、1024KB、2048KB。
  • 支援 OS-aware cost profile：normal、high_latency、high_memory_pressure，以及 runtime-derived custom profiles。
  • 輸出 CSV 結果，包含 DRAM traffic、estimated energy、estimated latency、peak SRAM usage、fusion depth 與 OS-aware total cost。
  • 從 simulation CSV 產生 summary CSV 與 SVG 圖表。
  • 提供 ROS2 FPS/QoS/overload 實驗腳本模板，預設為外部 MobileViT UAV pipeline。
  • 分析 ROS2 runtime CSV，輸出 latency、jitter、drop rate、deadline miss rate 與 SVG 圖表。
  • 在 explicit dropped rows 缺席時，可由 frame_id 缺口推估 dropped frames。
  • 將 ROS2 runtime summary 轉成 EDA-Lite 自訂 OS pressure profiles，包含 drop pressure。

目前仍不包含或尚未完整追溯：

  • raw ROS2 runtime CSVs under results/raw/ros2/.
  • results/csv/ros2_runtime_summary.csv.
  • 真實 MobileViT 權重。
  • ONNX parser。
  • 真實 LLM inference 或 KV-cache simulation。
  • 真實硬體 energy measurement。
  • 完整 EDA paper compiler reproduction。

Current checked-in ROS2 runtime-derived artifacts are documented in:

report/runtime_result_provenance.md

Check whether runtime artifacts are complete enough for measured-result reporting:

python3 scripts/check_runtime_provenance.py

🚀 完整端到端執行流程 (End-to-End Pipeline)

本專案所有的 .sh 腳本與 Python 模組已經高度自動化。請依照以下四個步驟順序執行，即可完整重現專題的所有數據與圖表：

步驟 0：環境建置 (Environment Setup)

本專案為了解決無 sudo 權限的伺服器環境，提供了基於 Conda 的 ROS 2 安裝腳本與依賴套件設定檔。請依序執行以下腳本完成環境初始化：

bash scripts/setup_env.sh
bash scripts/install_ros2_conda.sh

setup_env.sh creates or reuses the eda_lite conda environment. install_ros2_conda.sh should run after that environment exists.

步驟 1：執行 ROS 2 硬體壓力測試 (實機量測)

確保已設定好上述環境與 ROS 2 變數後，執行自動化總管腳本，它會自動限制 CPU/RAM 資源並連續跑完 FPS、QoS、Overload 三大測試，產出 22 份 CSV 原始數據：

bash scripts/run_baseline_experiments.sh

步驟 2：分析 ROS 2 數據與產生長條圖

將原始 CSV 進行彙整，推估系統 Jitter 與真實掉幀率，並產出視覺化長條圖：

python3 -m src.ros2_runtime.analyze_runtime \
  --input results/raw/ros2 \
  --summary-output results/csv/ros2_runtime_summary.csv \
  --figures-dir results/figures/ros2

步驟 3：轉換為 OS-Aware 懲罰特徵

將上一步的 ROS 2 掉幀與延遲數據，轉換為 EDA 模擬器看得懂的 JSON OS 懲罰設定檔：

python3 -m src.ros2_runtime.build_os_profiles \
  --input results/csv/ros2_runtime_summary.csv \
  --output experiments/eda_lite/runtime_os_profiles.json \
  --profile-prefix measured

(註：轉換完成後，請確保 experiments/eda_lite/mobilevit_runtime_aware_sweep.json 指向了該設定檔)

步驟 4：執行 EDA-Lite 模擬與產出最終 6 條線比較圖

依序執行 Pure (無 OS 懲罰) 與 OS-Aware (有 OS 懲罰) 的硬體模擬，最後使用客製化腳本將它們疊加成最終的報告對比圖：

bash scripts/rebuild_eda_lite_figures.sh

The script rebuilds MobileViT-S and MobileViT-XXS pure simulation CSVs, runtime-aware simulation CSVs, and the final six-line comparison figures.

完成後，最終報告圖表將全部安穩地躺在 results/figures/eda_lite/ 底下！

Repository Layout

OS_Final/
  AGENTS.md
  README.md
  跨平台輕量化模型效能分析與模擬.pdf
  src/
    eda_lite/
      model_loader.py
      schedulers.py
      cost_model.py
      run_simulation.py
      run_profile_sweep.py
      analyze_results.py
    ros2_runtime/
      analyze_runtime.py
      build_os_profiles.py
    ros2_yolo11/
      analyze_runtime.py
      detector_node_notes.md
  data/
    model_profiles/
      mobilevit_xxs_simplified.json
      mobilevit_s_simplified.json
      edge_cnn_simplified.json
      edge_cnn_large_simplified.json
      yolo11n_simplified.json
      yolo11s_simplified.json
  experiments/
    eda_lite/
      mobilevit_sram_sweep.json
      mobilevit_profile_sweep.json
      mobilevit_runtime_aware_sweep.json
      runtime_os_profiles.example.json
      edge_profile_sweep.json
      yolo11_profile_sweep.json
    ros2/
      run_fps_sweep.sh
      run_qos_sweep.sh
      run_overload_sweep.sh
      run_yolo11_comparison.sh
  results/
    csv/
    figures/
  report/
    ai_interaction_log.md
    final_report.md
    ros2_pipeline_integration.md
  tests/

Requirements

目前核心 simulator 與 analyzer 只使用 Python 標準函式庫，不需要額外安裝套件。

建議使用：

python3 --version

Run MobileViT Simulation

Single-profile SRAM sweep:

python3 -m src.eda_lite.run_simulation \
  --profile data/model_profiles/mobilevit_xxs_simplified.json \
  --config experiments/eda_lite/mobilevit_sram_sweep.json \
  --output results/csv/eda_lite_mobilevit_xxs_sram_sweep.csv

MobileViT-XXS vs MobileViT-S profile sweep:

python3 -m src.eda_lite.run_profile_sweep \
  --config experiments/eda_lite/mobilevit_profile_sweep.json \
  --output results/csv/eda_lite_mobilevit_profile_sweep.csv

Analyze MobileViT profile sweep:

python3 -m src.eda_lite.analyze_results \
  --input results/csv/eda_lite_mobilevit_profile_sweep.csv \
  --summary-output results/csv/eda_lite_mobilevit_profile_summary.csv \
  --figures-dir results/figures/mobilevit_s \
  --plot-os-profile normal \
  --plot-model mobilevit_s_simplified_profile_estimate

When a CSV contains multiple models, always pass --plot-model.

ROS2 Runtime Experiments

ROS2 scripts are stored in:

experiments/ros2/

This repository now includes a lightweight ROS2 MobileViT/UAV pipeline under src/ros2_uav_pipeline/. The scripts can launch the checked-in launch file directly, or use an installed ROS2 package if you set ROS2_LAUNCH_TARGET="".

Common environment variables:

export ROS2_WORKSPACE="$HOME/ros2-mobilevit-uav-pipeline"
export ROS2_SETUP="$ROS2_WORKSPACE/install/setup.bash"
export ROS2_PACKAGE="mobilevit_uav_pipeline"
export ROS2_LAUNCH_FILE="mobilevit_uav_pipeline.launch.py"
export ROS2_LAUNCH_TARGET="src/ros2_uav_pipeline/launch/mobilevit_uav_pipeline.launch.py"
export DATASET_PATH="$PWD/data/raw/images/coco128/images/train2017"
export RESULTS_DIR="$PWD/results/raw/ros2"
export MODEL_NAME="mobilevit_xxs"
export MODEL_PATH="mobilevit_xxs"

Prepare optional COCO128 image-folder input:

bash experiments/ros2/prepare_coco128.sh

COCO128 is used only as a compact image stream source for ROS2 runtime pressure tests. It is not a YOLO training or evaluation experiment. Dataset documentation: https://docs.ultralytics.com/datasets/detect/coco128/

Run the main sweeps:

bash experiments/ros2/run_fps_sweep.sh
bash experiments/ros2/run_qos_sweep.sh
bash experiments/ros2/run_overload_sweep.sh

By default, the scripts launch the checked-in launch file through ROS2_LAUNCH_TARGET. To launch an installed ROS2 package instead, set ROS2_LAUNCH_TARGET="" and use ROS2_PACKAGE plus ROS2_LAUNCH_FILE. The sweep scripts fail early if ros2 is not available, and require either GNU timeout or gtimeout so each scenario stops after RUN_SECONDS seconds.

Run the two-model baseline wrapper:

bash scripts/run_baseline_experiments.sh

On Linux, the wrapper uses taskset and systemd-run to apply 2-core / 2GB resource limits. On macOS or environments without those tools, it falls back to running the sweep scripts without Linux resource isolation; this is useful for functional checks but should not be reported as the constrained OS measurement condition.

The launch file is expected to accept arguments such as dataset_path, input_fps, qos_depth, model_path, model_name, fake_inference_delay_ms, deadline_ms, resource_profile, scenario, and output_csv. If the adapted ROS2 project uses different launch argument names, update the shell scripts.

Do not report runtime data as fully auditable measured results unless CSVs were generated by a real ROS2 run and the raw CSVs or runtime summary CSV are available.

The teammate pipeline integration checklist is documented in:

report/ros2_pipeline_integration.md

Analyze ROS2 Runtime CSV

After ROS2 experiments produce CSV files in results/raw/ros2, generate a runtime summary and figures:

python3 -m src.ros2_runtime.analyze_runtime \
  --input results/raw/ros2 \
  --summary-output results/csv/ros2_runtime_summary.csv \
  --figures-dir results/figures/ros2

Recommended ROS2 logger columns:

scenario,model_name,input_fps,qos_depth,fake_delay_ms,resource_profile,
frame_id,published_at_ns,received_at_ns,completed_at_ns,
latency_ms,inference_ms,deadline_miss,dropped,cpu_percent,memory_mb

Notes:

  • latency_ms may be provided directly.
  • If latency_ms is absent, the analyzer can derive latency from published_at_ns and completed_at_ns.
  • If explicit dropped rows are absent, the analyzer can infer dropped frames from monotonically increasing frame_id gaps.
  • If deadline_miss is absent or false and deadline_ms exists, the analyzer can infer deadline misses from latency.

Runtime summary columns:

  • scenario
  • model_name
  • input_fps
  • qos_depth
  • fake_delay_ms
  • resource_profile
  • total_rows
  • processed_frames
  • dropped_frames
  • drop_rate_pct
  • deadline_misses
  • deadline_miss_rate_pct
  • avg_latency_ms
  • p95_latency_ms
  • p99_latency_ms
  • jitter_ms
  • avg_inference_ms
  • avg_cpu_percent
  • avg_memory_mb

Build Runtime OS Profiles

Convert a ROS2 runtime summary into EDA-Lite OS pressure profiles:

python3 -m src.ros2_runtime.build_os_profiles \
  --input results/csv/ros2_runtime_summary.csv \
  --output experiments/eda_lite/runtime_os_profiles.json \
  --profile-prefix measured

Then update experiments/eda_lite/mobilevit_runtime_aware_sweep.json so os_profile_file points to runtime_os_profiles.json, and run:

python3 -m src.eda_lite.run_simulation \
  --profile data/model_profiles/mobilevit_xxs_simplified.json \
  --config experiments/eda_lite/mobilevit_runtime_aware_sweep.json \
  --output results/csv/eda_lite_mobilevit_runtime_aware_sweep.csv

mobilevit_runtime_aware_sweep.json uses runtime_os_profiles.example.json by default. That example file is not measured data. In the current branch, mobilevit_runtime_aware_sweep.json points to runtime_os_profiles.json, which contains teammate-provided runtime-derived measured_* profiles. See report/runtime_result_provenance.md before using those numbers in the final report.

Model Profile Format

模型 profile 使用簡化 JSON，不代表完整 ONNX graph。

必要欄位：

{
  "model_name": "example_model",
  "layers": [
    {
      "name": "conv1",
      "op_type": "Conv",
      "input_kb": 64,
      "output_kb": 48,
      "weight_kb": 48,
      "compute_macs": 120000000,
      "fusible": true
    }
  ]
}

Main profiles:

data/model_profiles/mobilevit_xxs_simplified.json
data/model_profiles/mobilevit_s_simplified.json

Legacy optional profiles:

data/model_profiles/edge_cnn_simplified.json
data/model_profiles/edge_cnn_large_simplified.json
data/model_profiles/yolo11n_simplified.json
data/model_profiles/yolo11s_simplified.json

Experiment Config Format

範例 config：

{
  "sram_kb": [256, 512, 1024, 2048],
  "schedulers": ["layerwise", "greedy", "dp"],
  "os_profiles": ["normal", "high_latency", "high_memory_pressure"],
  "cost_model": {
    "dram_energy_per_kb": 1.0,
    "compute_energy_per_mmac": 0.05,
    "dram_latency_per_kb_ms": 0.001,
    "compute_latency_per_mmac_ms": 0.01,
    "subgraph_overhead_ms": 0.02,
    "weight_tile_kb": 64.0,
    "alpha": 1.0,
    "beta": 100.0,
    "gamma": 50.0,
    "delta": 75.0,
    "deadline_ms": 33.3
  }
}

Main configs:

experiments/eda_lite/mobilevit_sram_sweep.json
experiments/eda_lite/mobilevit_profile_sweep.json
experiments/eda_lite/mobilevit_runtime_aware_sweep.json
experiments/eda_lite/runtime_os_profiles.example.json

Cost Model Notes

EDA-Lite v1 使用簡化成本模型：

os_aware_total_cost =
    estimated_energy
  + alpha * estimated_latency
  + beta  * deadline_miss_penalty
  + gamma * memory_pressure_penalty
  + delta * communication_drop_penalty

其中：

  • DRAM traffic 會計入 subgraph 的第一個 input、最後一個 output 與所有 weights。
  • 融合後的中間 activation read/write 會被省略，用來模擬 layer fusion 減少 DRAM traffic。
  • SRAM usage 以 input、output、最大中間 activation 與 weight tile 估計。
  • drop_rate 由 ROS2 runtime summary 的 drop_rate_pct 轉換而來，用來表示 DDS/UDP/接收端壓力造成的 communication pressure。
  • OS profile 可使用內建 coarse pressure profile，也可由 ROS2 runtime summary 轉成自訂 profile。

OS profile JSON shape:

{
  "os_profiles": ["measured_example"],
  "os_profile_definitions": {
    "measured_example": {
      "latency_multiplier": 1.0,
      "deadline_miss_rate": 0.02,
      "memory_pressure": 0.45,
      "drop_rate": 0.03
    }
  }
}

drop_rate is optional for backward compatibility and defaults to 0.0.

Run Tests

在專案根目錄執行：

python3 -m unittest discover -s tests

目前測試會檢查：

  • simulator 可以讀取 JSON model profile。
  • MobileViT profiles and profile sweep configs can run.
  • layerwise、greedy、dp scheduling behavior.
  • analysis summary and SVG generation.
  • multi-model analysis requires --plot-model.
  • ROS2 runtime analyzer can derive latency from timestamps.
  • ROS2 runtime analyzer can infer dropped frames from frame_id gaps.
  • ROS2 runtime summary can convert deadline, memory, and drop pressure into OS profiles.
  • old OS profile JSON without drop_rate remains compatible.

Reporting Rules

  • Clearly separate measured ROS2 runtime results from EDA-Lite simulation estimates.
  • Label EDA-Lite energy and latency as estimates or simulation outputs.
  • Do not present estimated energy as real hardware energy.
  • Do not claim this project fully reproduces the original EDA compiler.
  • Explain the simplified memory, energy, MobileViT block, and communication-drop assumptions.
  • Explain that MobileViT profiles are coarse simulation profiles, not exact ONNX model structures.
  • If using runtime_os_profiles.example.json, state that it is an example profile, not measured data.
  • Only call ROS2 data "measured" after it comes from actual ROS2 experiment CSVs.
  • YOLO-related content is paused; do not frame the project as a YOLO project unless the direction explicitly changes.

Development Notes

  • 每新增一個可執行功能，同步更新本 README。
  • 保持 EDA-Lite simulator deterministic，方便報告解釋與測試。
  • 清楚區分 runtime measurement 與 simulation estimate。
  • 不宣稱完整重現 EDA paper，只稱為 EDA-inspired lightweight simulator。
  • ROS2 CSV 只有實際跑過 pipeline 後才能標示為 runtime measurement；synthetic 或測試資料不得當作實測結果。

下一階段建議：

  1. 撰寫 report methodology 與 limitations。
  2. 實際 ROS2 workspace 準備完成後，執行 FPS/QoS/overload experiments 並保存 raw CSV。
  3. 用真實 ROS2 summary 產生 runtime_os_profiles.json，再重跑 mobilevit_runtime_aware_sweep.json。
  4. 產生 final figures 並寫 results discussion。

About

No description, website, or topics provided.

Resources

Readme
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Contributors

Languages