| # 并行压力测试数据 |
|
|
| 每 GPU 单独保存 `results.json`。 |
|
|
| ## 文件 |
|
|
| | 文件 | 内容 | |
| |------|------| |
| | `RTX4090_24GB/results.json` | 4090 压力测试（TanhNormal, 1–12 并发, 5K 步/实验, warmup） | |
| | `RTX4090_24GB/results_ds_stereographic.json` | 4090 压力测试（DS stereographic, 1–10 并发, 10K 步/实验, 无 warmup, schedule.py 调度） | |
| | `RTX6000Ada_48GB/results.json` | 6000 Ada 压力测试（1–10 并发，20K 步/实验） | |
|
|
| ## results.json 格式 |
|
|
| ```json |
| { |
| "gpu": "NVIDIA GeForce RTX 4090", |
| "vram": "24GB", |
| "single_baseline_it_s": 228, |
| "mem_fraction": 0.05, |
| "steps_per_test": 5000, |
| "cores_total": 32, |
| "results": [ |
| { |
| "concurrency": 1, |
| "cores_per": 32, |
| "total_it_s": 228, |
| "scale": 1.0, |
| "per_proc": 228, |
| "gpu_util": "33%", |
| "vram_used": "2GB", |
| "status": "OK" |
| } |
| ] |
| } |
| ``` |
|
|
| ### 字段 |
|
|
| | 字段 | 含义 | |
| |------|------| |
| | `concurrency` | 并行实验数 | |
| | `cores_per` | 每进程 CPU 核数 | |
| | `total_it_s` | 全进程合计吞吐（it/s） | |
| | `scale` | 相对单进程的倍率 | |
| | `per_proc` | 每进程平均吞吐 | |
| | `gpu_util` | GPU 利用率（nvidia-smi） | |
| | `vram_used` | 显存占用 | |
| | `status` | OK / CRASHED | |
|
|
| ### raw_measurements 字段（results.json 内） |
| |
| `raw_measurements` 按并发数分组，记录每个进程的实测吞吐： |
|
|
| ```json |
| { |
| "8": { |
| "total_launched": 8, |
| "completed": 8, |
| "per_process_it_s": [66.2, 68.1, 67.5, 67.3, 59.4, 58.8, 58.6, 63.1], |
| "total_it_s": 504.8, |
| "mean_per_proc": 63.1 |
| } |
| } |
| ``` |
|
|
| | 字段 | 含义 | |
| |------|------| |
| | `total_launched` | 启动进程数 | |
| | `completed` | 成功完成数 | |
| | `per_process_it_s` | 每个进程的稳态吞吐 | |
| | `total_it_s` | 合计吞吐 | |
| | `mean_per_proc` | 进程平均吞吐 | |
|
|
| ## 实验信息 |
|
|
| `results.json` 顶层包含实验参数： |
|
|
| | 字段 | 4090 | 6000 Ada | |
| |------|:--:|:--:| |
| | `experiment` | 并行训练压力测试 | 并行训练压力测试 | |
| | `method` | RLPD (TanhNormal) | RLPD (TanhNormal) | |
| | `env` | cube-triple-play-singletask-task2-v0 | cube-triple-play-singletask-task2-v0 | |
| | `entry` | main_online.py | main_online.py | |
| | `horizon_length` | 1 | 1 | |
| | `action_chunking` | False | False | |
| | `online_steps` | 5000 | 20000 | |
| | `eval_interval` | 0 | 0 | |
| | `cores_total` | 32 | 32 | |
|
|
