GPU测试

2026-01-20 10:54:30 +08:00
parent 8463f5a571
commit 8e9de9c858
59 changed files with 18934 additions and 0 deletions
--- a/benchmark/visualizer.py
+++ b/benchmark/visualizer.py
@@ -0,0 +1,315 @@
+"""
+Benchmark 结果可视化模块
+"""
+
+import os
+from typing import List, Dict, Any
+from datetime import datetime
+
+import numpy as np
+
+from .results import TestResult
+from .utils import ensure_dir, setup_logging
+
+logger = setup_logging()
+
+
+def generate_visualizations(results: List[TestResult], output_dir: str) -> List[str]:
+    """生成所有可视化图表"""
+    try:
+        import matplotlib.pyplot as plt
+        import matplotlib
+        matplotlib.use('Agg')  # 非交互式后端
+        plt.rcParams['font.sans-serif'] = ['SimHei', 'DejaVu Sans']
+        plt.rcParams['axes.unicode_minus'] = False
+    except ImportError:
+        logger.warning("matplotlib 未安装，跳过可视化")
+        return []
+    
+    ensure_dir(output_dir)
+    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+    generated_files = []
+    
+    # 1. GPU 利用率 vs 摄像头数量
+    fig_path = _plot_gpu_vs_cameras(results, output_dir, timestamp)
+    if fig_path:
+        generated_files.append(fig_path)
+    
+    # 2. 吞吐量 vs Batch Size
+    fig_path = _plot_throughput_vs_batch(results, output_dir, timestamp)
+    if fig_path:
+        generated_files.append(fig_path)
+    
+    # 3. 延迟分布
+    fig_path = _plot_latency_distribution(results, output_dir, timestamp)
+    if fig_path:
+        generated_files.append(fig_path)
+    
+    # 4. 实时性热力图
+    fig_path = _plot_realtime_heatmap(results, output_dir, timestamp)
+    if fig_path:
+        generated_files.append(fig_path)
+    
+    # 5. GPU 饱和点分析
+    fig_path = _plot_saturation_analysis(results, output_dir, timestamp)
+    if fig_path:
+        generated_files.append(fig_path)
+    
+    logger.info(f"生成 {len(generated_files)} 个可视化图表")
+    return generated_files
+
+
+def _plot_gpu_vs_cameras(results: List[TestResult], output_dir: str, timestamp: str) -> str:
+    """GPU 利用率 vs 摄像头数量"""
+    import matplotlib.pyplot as plt
+    
+    fig, ax = plt.subplots(figsize=(10, 6))
+    
+    # 按 batch size 分组
+    batch_sizes = sorted(set(r.batch_size for r in results))
+    camera_counts = sorted(set(r.num_cameras for r in results))
+    
+    for bs in batch_sizes:
+        bs_results = [r for r in results if r.batch_size == bs]
+        cameras = []
+        gpu_utils = []
+        for cam in camera_counts:
+            cam_results = [r for r in bs_results if r.num_cameras == cam]
+            if cam_results:
+                cameras.append(cam)
+                gpu_utils.append(np.mean([r.gpu_utilization_avg for r in cam_results]))
+        
+        if cameras:
+            ax.plot(cameras, gpu_utils, marker='o', label=f'Batch={bs}')
+    
+    ax.axhline(y=85, color='r', linestyle='--', label='饱和阈值 (85%)')
+    ax.set_xlabel('摄像头数量')
+    ax.set_ylabel('GPU 利用率 (%)')
+    ax.set_title('GPU 利用率 vs 摄像头数量')
+    ax.legend()
+    ax.grid(True, alpha=0.3)
+    ax.set_ylim(0, 100)
+    
+    fig_path = os.path.join(output_dir, f'gpu_vs_cameras_{timestamp}.png')
+    plt.savefig(fig_path, dpi=150, bbox_inches='tight')
+    plt.close()
+    
+    return fig_path
+
+
+def _plot_throughput_vs_batch(results: List[TestResult], output_dir: str, timestamp: str) -> str:
+    """吞吐量 vs Batch Size"""
+    import matplotlib.pyplot as plt
+    
+    fig, ax = plt.subplots(figsize=(10, 6))
+    
+    batch_sizes = sorted(set(r.batch_size for r in results))
+    camera_counts = sorted(set(r.num_cameras for r in results))
+    
+    for cam in camera_counts:
+        cam_results = [r for r in results if r.num_cameras == cam]
+        batches = []
+        throughputs = []
+        for bs in batch_sizes:
+            bs_results = [r for r in cam_results if r.batch_size == bs]
+            if bs_results:
+                batches.append(bs)
+                throughputs.append(np.mean([r.total_throughput_fps for r in bs_results]))
+        
+        if batches:
+            ax.plot(batches, throughputs, marker='s', label=f'{cam} 路摄像头')
+    
+    ax.set_xlabel('Batch Size')
+    ax.set_ylabel('总吞吐量 (FPS)')
+    ax.set_title('吞吐量 vs Batch Size')
+    ax.legend()
+    ax.grid(True, alpha=0.3)
+    ax.set_xticks(batch_sizes)
+    
+    fig_path = os.path.join(output_dir, f'throughput_vs_batch_{timestamp}.png')
+    plt.savefig(fig_path, dpi=150, bbox_inches='tight')
+    plt.close()
+    
+    return fig_path
+
+
+def _plot_latency_distribution(results: List[TestResult], output_dir: str, timestamp: str) -> str:
+    """延迟分布对比"""
+    import matplotlib.pyplot as plt
+    
+    fig, axes = plt.subplots(1, 2, figsize=(14, 5))
+    
+    batch_sizes = sorted(set(r.batch_size for r in results))
+    
+    # 左图: 平均延迟
+    ax1 = axes[0]
+    x = np.arange(len(batch_sizes))
+    width = 0.35
+    
+    avg_latencies = [np.mean([r.avg_latency_ms for r in results if r.batch_size == bs]) for bs in batch_sizes]
+    p95_latencies = [np.mean([r.p95_latency_ms for r in results if r.batch_size == bs]) for bs in batch_sizes]
+    
+    ax1.bar(x - width/2, avg_latencies, width, label='平均延迟')
+    ax1.bar(x + width/2, p95_latencies, width, label='P95 延迟')
+    ax1.set_xlabel('Batch Size')
+    ax1.set_ylabel('延迟 (ms)')
+    ax1.set_title('延迟 vs Batch Size')
+    ax1.set_xticks(x)
+    ax1.set_xticklabels(batch_sizes)
+    ax1.legend()
+    ax1.grid(True, alpha=0.3, axis='y')
+    
+    # 右图: 按摄像头数量
+    ax2 = axes[1]
+    camera_counts = sorted(set(r.num_cameras for r in results))
+    
+    avg_by_cam = [np.mean([r.avg_latency_ms for r in results if r.num_cameras == cam]) for cam in camera_counts]
+    p95_by_cam = [np.mean([r.p95_latency_ms for r in results if r.num_cameras == cam]) for cam in camera_counts]
+    
+    x2 = np.arange(len(camera_counts))
+    ax2.bar(x2 - width/2, avg_by_cam, width, label='平均延迟')
+    ax2.bar(x2 + width/2, p95_by_cam, width, label='P95 延迟')
+    ax2.set_xlabel('摄像头数量')
+    ax2.set_ylabel('延迟 (ms)')
+    ax2.set_title('延迟 vs 摄像头数量')
+    ax2.set_xticks(x2)
+    ax2.set_xticklabels(camera_counts)
+    ax2.legend()
+    ax2.grid(True, alpha=0.3, axis='y')
+    
+    plt.tight_layout()
+    fig_path = os.path.join(output_dir, f'latency_distribution_{timestamp}.png')
+    plt.savefig(fig_path, dpi=150, bbox_inches='tight')
+    plt.close()
+    
+    return fig_path
+
+
+def _plot_realtime_heatmap(results: List[TestResult], output_dir: str, timestamp: str) -> str:
+    """实时性热力图"""
+    import matplotlib.pyplot as plt
+    
+    batch_sizes = sorted(set(r.batch_size for r in results))
+    camera_counts = sorted(set(r.num_cameras for r in results))
+    
+    # 创建矩阵
+    matrix = np.zeros((len(camera_counts), len(batch_sizes)))
+    
+    for i, cam in enumerate(camera_counts):
+        for j, bs in enumerate(batch_sizes):
+            matching = [r for r in results if r.num_cameras == cam and r.batch_size == bs]
+            if matching:
+                # 实时性比例
+                realtime_ratio = sum(1 for r in matching if r.is_realtime_capable) / len(matching)
+                matrix[i, j] = realtime_ratio * 100
+    
+    fig, ax = plt.subplots(figsize=(10, 8))
+    
+    im = ax.imshow(matrix, cmap='RdYlGn', aspect='auto', vmin=0, vmax=100)
+    
+    ax.set_xticks(np.arange(len(batch_sizes)))
+    ax.set_yticks(np.arange(len(camera_counts)))
+    ax.set_xticklabels(batch_sizes)
+    ax.set_yticklabels(camera_counts)
+    ax.set_xlabel('Batch Size')
+    ax.set_ylabel('摄像头数量')
+    ax.set_title('实时性热力图 (绿色=可实时, 红色=不可实时)')
+    
+    # 添加数值标注
+    for i in range(len(camera_counts)):
+        for j in range(len(batch_sizes)):
+            text = ax.text(j, i, f'{matrix[i, j]:.0f}%',
+                          ha='center', va='center', color='black', fontsize=9)
+    
+    plt.colorbar(im, ax=ax, label='实时性比例 (%)')
+    
+    fig_path = os.path.join(output_dir, f'realtime_heatmap_{timestamp}.png')
+    plt.savefig(fig_path, dpi=150, bbox_inches='tight')
+    plt.close()
+    
+    return fig_path
+
+
+def _plot_saturation_analysis(results: List[TestResult], output_dir: str, timestamp: str) -> str:
+    """GPU 饱和点分析"""
+    import matplotlib.pyplot as plt
+    
+    fig, axes = plt.subplots(2, 2, figsize=(14, 10))
+    
+    camera_counts = sorted(set(r.num_cameras for r in results))
+    batch_sizes = sorted(set(r.batch_size for r in results))
+    
+    # 1. GPU 利用率 vs 吞吐量
+    ax1 = axes[0, 0]
+    for bs in batch_sizes:
+        bs_results = [r for r in results if r.batch_size == bs]
+        throughputs = [r.total_throughput_fps for r in bs_results]
+        gpu_utils = [r.gpu_utilization_avg for r in bs_results]
+        ax1.scatter(throughputs, gpu_utils, label=f'Batch={bs}', alpha=0.7)
+    
+    ax1.axhline(y=85, color='r', linestyle='--', alpha=0.5)
+    ax1.set_xlabel('吞吐量 (FPS)')
+    ax1.set_ylabel('GPU 利用率 (%)')
+    ax1.set_title('GPU 利用率 vs 吞吐量')
+    ax1.legend()
+    ax1.grid(True, alpha=0.3)
+    
+    # 2. 显存使用
+    ax2 = axes[0, 1]
+    for bs in batch_sizes:
+        bs_results = [r for r in results if r.batch_size == bs]
+        cameras = []
+        mem_used = []
+        for cam in camera_counts:
+            cam_results = [r for r in bs_results if r.num_cameras == cam]
+            if cam_results:
+                cameras.append(cam)
+                mem_used.append(np.mean([r.memory_used_mb for r in cam_results]))
+        if cameras:
+            ax2.plot(cameras, mem_used, marker='o', label=f'Batch={bs}')
+    
+    ax2.set_xlabel('摄像头数量')
+    ax2.set_ylabel('显存使用 (MB)')
+    ax2.set_title('显存使用 vs 摄像头数量')
+    ax2.legend()
+    ax2.grid(True, alpha=0.3)
+    
+    # 3. 丢帧率
+    ax3 = axes[1, 0]
+    for bs in batch_sizes:
+        bs_results = [r for r in results if r.batch_size == bs]
+        cameras = []
+        drop_rates = []
+        for cam in camera_counts:
+            cam_results = [r for r in bs_results if r.num_cameras == cam]
+            if cam_results:
+                cameras.append(cam)
+                drop_rates.append(np.mean([r.frame_drop_rate for r in cam_results]))
+        if cameras:
+            ax3.plot(cameras, drop_rates, marker='s', label=f'Batch={bs}')
+    
+    ax3.axhline(y=5, color='r', linestyle='--', label='阈值 (5%)')
+    ax3.set_xlabel('摄像头数量')
+    ax3.set_ylabel('丢帧率 (%)')
+    ax3.set_title('丢帧率 vs 摄像头数量')
+    ax3.legend()
+    ax3.grid(True, alpha=0.3)
+    
+    # 4. 饱和状态统计
+    ax4 = axes[1, 1]
+    saturated_count = sum(1 for r in results if r.is_gpu_saturated)
+    not_saturated_count = len(results) - saturated_count
+    
+    ax4.pie([not_saturated_count, saturated_count], 
+            labels=['未饱和', '已饱和'],
+            colors=['#2ecc71', '#e74c3c'],
+            autopct='%1.1f%%',
+            startangle=90)
+    ax4.set_title(f'GPU 饱和状态分布\n(共 {len(results)} 个测试)')
+    
+    plt.tight_layout()
+    fig_path = os.path.join(output_dir, f'saturation_analysis_{timestamp}.png')
+    plt.savefig(fig_path, dpi=150, bbox_inches='tight')
+    plt.close()
+    
+    return fig_path