import os
import json
import requests
from flask import Flask, jsonify, request, render_template
from datetime import datetime
import logging
import glob
import time
import argparse
import re
app = Flask(__name__)
# 配置日志
logging.basicConfig(level=logging.DEBUG, format='%(asctime)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)
# 定义基础路径
BASE_PATH = r"C:\Users\l30078648\Desktop\250730"
DOWNLOAD_PATH = os.path.join(BASE_PATH, "downloads")
# S3 OBS API配置
S3_OBS_API_URL = "https://adsci.api.ias.huawei.com/icollect/get_s3_obs_file_download_url/"
S3_OBS_HEADERS = {
"Accept-Language": "zh-CN",
"Content-Type": "application/json"
}
USER_ID = "h00512356"
# 芯片型号到文件后缀的映射
CHIP_FILENAME_MAP = {
"BS9SX1A": "BS9SX1AA", # 添加额外的'A'后缀
"Ascend610Lite": "Ascend610Lite"
}
# 定义芯片和模型组路径映射
TRACK_PATHS = {
"Ascend610Lite": {
"rl_nn": os.path.join(BASE_PATH, ".track", "Ascend610Lite", "rl_nn"),
"rsc_nn": os.path.join(BASE_PATH, ".track", "Ascend610Lite", "rsc_nn"),
"prediction_nn": os.path.join(BASE_PATH, ".track", "Ascend610Lite", "prediction_nn")
},
"BS9SX1A": {
"rl_nn": os.path.join(BASE_PATH, ".track", "BS9SX1A", "rl_nn"),
"rsc_nn": os.path.join(BASE_PATH, ".track", "BS9SX1A", "rsc_nn"),
"prediction_nn": os.path.join(BASE_PATH, ".track", "BS9SX1A", "prediction_nn")
}
}
JSON_PATHS = {
"Ascend610Lite": {
"rl_nn": os.path.join(DOWNLOAD_PATH, "Ascend610Lite", "rl_nn"),
"rsc_nn": os.path.join(DOWNLOAD_PATH, "Ascend610Lite", "rsc_nn"),
"prediction_nn": os.path.join(DOWNLOAD_PATH, "Ascend610Lite", "prediction_nn")
},
"BS9SX1A": {
"rl_nn": os.path.join(DOWNLOAD_PATH, "BS9SX1A", "rl_nn"),
"rsc_nn": os.path.join(DOWNLOAD_PATH, "BS9SX1A", "rsc_nn"),
"prediction_nn": os.path.join(DOWNLOAD_PATH, "BS9SX1A", "prediction_nn")
}
}
def send_request(method, url, headers=None, data=None, timeout=30, max_try_times=1, **kwargs):
"""发送HTTP请求（带重试机制）"""
try_times = 0
response = None
if data is None:
data = {}
if 'files' in kwargs:
request_body = data
else:
request_body = json.dumps(data) if not isinstance(data, str) else data.encode("utf-8")
while try_times < max_try_times:
try:
response = requests.request(
method=method,
url=url,
headers=headers,
data=request_body,
timeout=timeout,
verify=False,
**kwargs
)
if response.status_code in (200, 201):
break
except requests.exceptions.Timeout:
logger.warning(f"请求超时: url={url}")
except Exception as e:
logger.exception(f"请求失败: url={url}, 错误: {str(e)}")
if response:
logger.error(f"响应内容: {response.text}")
try_times += 1
logger.warning(f"请求失败, url={url}, 重试中 ({try_times}/{max_try_times})")
if max_try_times > 1:
time.sleep(5)
return response
def extract_build_ids_from_track(chip, group):
"""从.track文件夹中提取build_id和完整模型名称数据，并生成OBS链接"""
build_data = {}
build_missing = []
if chip not in TRACK_PATHS or group not in TRACK_PATHS[chip]:
logger.error(f"无效路径: {chip}/{group}")
return build_data, build_missing
group_path = TRACK_PATHS[chip][group]
if not os.path.exists(group_path):
logger.error(f"原始路径不存在: {group_path}")
return build_data, build_missing
# 扫描所有JSON文件
json_files = glob.glob(os.path.join(group_path, "*.json"))
logger.info(f"在路径 {group_path} 中找到 {len(json_files)} 个JSON文件")
for json_file in json_files:
try:
with open(json_file, 'r', encoding='utf-8') as f:
data = json.load(f)
# 获取完整模型名称（带后缀）
model_full_name = os.path.splitext(os.path.basename(json_file))[0]
# 提取build_id
build_id = data.get('build_id')
# 处理不同格式的ID
if isinstance(build_id, int):
build_id = str(build_id)
# 生成OBS链接
obs_url = None
if build_id:
obs_url = f"https://adsci.ias.huawei.com/icollect/show_obs_objects/?s3_obs_url=hpct/nn_model_prebuild/{build_id}_PREBUILD_OUTPUTS"
# 存储build_id、模型全名和OBS链接
build_data[model_full_name] = {
"build_id": build_id,
"model_full_name": model_full_name,
"obs_url": obs_url # 新增OBS链接字段
}
if build_id:
logger.debug(f"提取成功: {model_full_name} -> build_id: {build_id}, OBS链接: {obs_url}")
else:
logger.warning(f"文件 {json_file} 中没有找到build_id字段")
build_missing.append(model_full_name)
except Exception as e:
logger.error(f"解析原始文件 {json_file} 时出错: {str(e)}")
build_missing.append(os.path.basename(json_file))
continue
logger.info(f"成功提取 {len(build_data) - len(build_missing)} 个build_id")
return build_data, build_missing
def get_chip_filename(chip):
"""获取芯片对应的正确文件名"""
# 使用映射表获取正确的文件名后缀
return CHIP_FILENAME_MAP.get(chip, f"{chip}")
def download_file_with_build_id(model_full_name, chip, group, build_id):
"""使用build_id下载JSON文件"""
if not build_id:
logger.warning(f"跳过缺少build_id的文件: {model_full_name}")
return None, None
# 获取芯片对应的正确文件名
chip_filename = get_chip_filename(chip)
# 构建S3 OBS路径
s3_obs_url = f"hpct/nn_model_prebuild/{build_id}_PREBUILD_OUTPUTS/profiling_summary/"
file_name = f"{chip_filename}_prof_record.json"
# 获取下载URL
download_url, error = get_download_url(build_id, s3_obs_url, file_name)
if not download_url:
logger.error(f"获取下载URL失败: {error}")
return None, None
# 下载并保存文件
save_path, file_data = download_and_save(download_url, model_full_name, chip, group, file_name)
return save_path, file_data
def get_download_url(build_id, s3_obs_url, file_name):
"""获取S3 OBS下载URL"""
payload = {
"user_id": USER_ID,
"s3_obs_url": s3_obs_url,
"file_name": file_name
}
logger.info(f"请求下载URL: {payload}")
response = send_request(
"POST",
S3_OBS_API_URL,
headers=S3_OBS_HEADERS,
data=payload
)
if not response or response.status_code != 200:
logger.error(f"获取下载URL失败: 状态码={getattr(response, 'status_code', '无响应')}")
return None, "下载URL获取失败"
try:
result = response.json()
download_url = result.get("download_url")
if not download_url:
logger.error(f"响应中缺少download_url字段: {result}")
return None, "响应中缺少download_url字段"
# 修复URL端口问题：确保使用10001端口
fixed_url = fix_download_url_port(download_url)
return fixed_url, None
except Exception as e:
logger.error(f"解析响应失败: {str(e)}")
return None, f"解析响应失败: {str(e)}"
def fix_download_url_port(url):
"""修正下载URL的端口号"""
# 使用正则表达式匹配并替换端口号
pattern = r"(\w+\.huawei\.com):\d+"
replacement = r"\1:10001"
fixed_url = re.sub(pattern, replacement, url)
if fixed_url != url:
logger.info(f"URL端口已修正: {url} -> {fixed_url}")
return fixed_url
def download_and_save(download_url, model_full_name, chip, group, file_name):
"""下载文件并保存到本地"""
logger.info(f"下载文件: {download_url}")
file_response = send_request("GET", download_url, timeout=60)
if not file_response or file_response.status_code != 200:
logger.error(f"文件下载失败: 状态码={getattr(file_response, 'status_code', '无响应')}")
return None, None
# 保存文件到本地
save_dir = JSON_PATHS[chip][group]
os.makedirs(save_dir, exist_ok=True)
save_file_name = f"{model_full_name}_{file_name}"
save_path = os.path.join(save_dir, save_file_name)
try:
with open(save_path, 'wb') as f:
f.write(file_response.content)
logger.info(f"文件保存成功: {save_path}")
# 尝试解析文件内容
try:
with open(save_path, 'r', encoding='utf-8') as f:
file_data = json.load(f)
return save_path, file_data
except Exception as e:
logger.error(f"解析下载的JSON文件失败: {save_path}, 错误: {str(e)}")
return save_path, None
except Exception as e:
logger.error(f"保存文件失败: {save_path}, 错误: {str(e)}")
return None, None
def download_files_with_build_ids(build_id_map, chip, group):
"""使用提取的build_id下载所有文件"""
downloaded_files = []
download_errors = []
for model_full_name, build_info in build_id_map.items():
build_id = build_info["build_id"]
if not build_id:
logger.warning(f"跳过缺少build_id的文件: {model_full_name}")
downloaded_files.append({
"model": model_full_name,
"status": "skipped",
"reason": "missing_build_id"
})
continue
logger.info(f"下载文件: {model_full_name} (build_id={build_id})")
save_path, file_data = download_file_with_build_id(model_full_name, chip, group, build_id)
if save_path:
if file_data:
status = "success"
else:
status = "partial_success"
download_errors.append({
"model": model_full_name,
"error": "文件解析失败"
})
downloaded_files.append({
"model": model_full_name,
"status": status
})
else:
logger.warning(f"下载失败: {model_full_name}")
download_errors.append({
"model": model_full_name,
"error": "下载失败"
})
downloaded_files.append({
"model": model_full_name,
"status": "failed"
})
return downloaded_files, download_errors
def extract_om_and_mean_ddr(data):
"""
从JSON数据中提取.om后缀键的值和mean_ddr值
:param data: JSON数据字典
:return: 元组 (om_value, mean_ddr_value, bandwidth_value)
"""
om_value = None
mean_ddr = None
bandwidth = None
# 查找所有以.om结尾的键并提取第一个值
om_keys = [key for key in data.keys() if key.endswith('.om')]
if om_keys:
om_value = data[om_keys[0]]
# 查找mean_ddr值（直接查找或嵌套查找）
if 'mean_ddr' in data:
mean_ddr = data['mean_ddr']
else:
# 在嵌套结构中查找mean_ddr
for value in data.values():
if isinstance(value, dict) and 'mean_ddr' in value:
mean_ddr = value['mean_ddr']
break
# 计算带宽值 (带宽 = 时延 * mean_ddr * 10 / 1000)
if om_value is not None and mean_ddr is not None:
try:
bandwidth = (float(om_value) * float(mean_ddr) * 10) / 1000
bandwidth = round(bandwidth, 4) # 保留4位小数
except (TypeError, ValueError):
logger.error(f"带宽计算失败: om_value={om_value}, mean_ddr={mean_ddr}")
return om_value, mean_ddr, bandwidth # 返回三个值
def parse_json_file(file_path, file_data=None):
"""解析JSON文件，提取性能数据和完整模型名称"""
if file_data is None:
try:
with open(file_path, 'r', encoding='utf-8') as f:
file_data = json.load(f)
except Exception as e:
logger.error(f"读取文件出错 {file_path}: {e}")
return None
# 从文件名中提取完整模型名称（带后缀）
file_name = os.path.basename(file_path)
# 文件名格式: "模型全名_芯片_prof_record.json"
# 确保与.track文件夹中的名称完全匹配
model_full_name = file_name.rsplit('_', 2)[0] # 去掉最后两个部分
# 提取.om值、mean_ddr值和带宽值
om_value, mean_ddr, bandwidth = extract_om_and_mean_ddr(file_data)
# 获取文件元数据
file_size = os.path.getsize(file_path)
last_modified = datetime.fromtimestamp(os.path.getmtime(file_path)).isoformat()
return {
"model_full_name": model_full_name,
"file_name": file_name,
"file_path": file_path,
"file_size": file_size,
"last_modified": last_modified,
"om_value": om_value,
"mean_ddr": mean_ddr,
"bandwidth": bandwidth # 新增带宽字段
}
def print_all_om_and_mean_ddr():
"""打印所有设备/组的所有.om、mean_ddr和带宽值到终端"""
print("\n" + "=" * 100)
print("JSON性能数据提取报告 - 时延(ms)、DDR和带宽(Mbps)")
print("=" * 100)
# 获取所有设备目录
devices = [d for d in os.listdir(DOWNLOAD_PATH)
if os.path.isdir(os.path.join(DOWNLOAD_PATH, d))]
if not devices:
print(f"在 {DOWNLOAD_PATH} 中未找到设备目录")
return
total_files = 0
valid_files = 0
for device in devices:
device_path = os.path.join(DOWNLOAD_PATH, device)
# 获取设备下的所有组目录
groups = [g for g in os.listdir(device_path)
if os.path.isdir(os.path.join(device_path, g))]
if not groups:
print(f"\n设备 {device} 中没有模型组")
continue
print(f"\n设备: {device} ({len(groups)}个模型组)")
print("-" * 90)
for group in groups:
group_path = os.path.join(device_path, group)
json_files = glob.glob(os.path.join(group_path, "*.json"))
if not json_files:
print(f" ├── 组 {group}: 没有JSON文件")
continue
print(f" ├── 组 {group}: {len(json_files)}个模型")
# 打印组内所有模型的指标
for json_file in json_files:
metrics = parse_json_file(json_file)
filename = os.path.basename(json_file)
om_str = f"{metrics['om_value']:.4f}" if metrics['om_value'] is not None else "N/A"
ddr_str = f"{metrics['mean_ddr']:.4f}" if metrics['mean_ddr'] is not None else "N/A"
bw_str = f"{metrics['bandwidth']:.4f}" if metrics['bandwidth'] is not None else "N/A"
print(f" │ ├── {metrics['model_full_name']} | 时延: {om_str:<8} | DDR: {ddr_str:<8} | 带宽: {bw_str:<8}")
total_files += 1
if metrics['om_value'] is not None and metrics['mean_ddr'] is not None:
valid_files += 1
print("\n" + "=" * 100)
print(f"扫描完成: 共处理 {total_files} 个文件, 有效数据 {valid_files} 个 ({valid_files/total_files*100:.1f}%)")
print("=" * 100 + "\n")
def get_performance_data(chip, group, refresh_data=False):
"""获取性能数据，refresh_data控制是否下载新文件"""
performance_data = {
"status": "success",
"models": [],
"timestamp": datetime.now().isoformat(),
"chip_type": chip,
"group": group,
"json_path": JSON_PATHS[chip].get(group, "") if chip in JSON_PATHS else "",
"track_path": TRACK_PATHS[chip].get(group, "") if chip in TRACK_PATHS else "",
"file_count": 0,
"build_id_count": 0,
"refresh_performed": refresh_data,
"download_errors": [],
"downloaded_files": [],
"build_missing": []
}
# 1. 从.track文件夹提取build_id和模型全名
build_id_map, build_missing = extract_build_ids_from_track(chip, group)
performance_data["build_id_count"] = len(build_id_map) - len(build_missing)
performance_data["build_missing"] = build_missing
# 2. 检查路径有效性
if chip not in JSON_PATHS or group not in JSON_PATHS[chip]:
performance_data["status"] = "error"
performance_data["error"] = "无效芯片或模型组"
return performance_data
group_path = JSON_PATHS[chip][group]
if not os.path.exists(group_path):
performance_data["status"] = "error"
performance_data["error"] = "JSON路径不存在"
return performance_data
# 确保目录存在
os.makedirs(group_path, exist_ok=True)
# 3. 刷新时下载所有文件
if refresh_data:
logger.info(f"开始刷新下载，共有{len(build_id_map)}个文件需要处理")
downloaded_files, download_errors = download_files_with_build_ids(build_id_map, chip, group)
performance_data["downloaded_files"] = downloaded_files
performance_data["download_errors"] = download_errors
performance_data["refresh_status"] = "executed"
else:
performance_data["refresh_status"] = "skipped"
logger.info("跳过文件下载，仅使用本地现有数据")
# 4. 处理所有JSON文件并提取指标
json_files = glob.glob(os.path.join(group_path, "*.json"))
performance_data["file_count"] = len(json_files)
for json_file in json_files:
model_data_parsed = parse_json_file(json_file)
if model_data_parsed:
model_full_name = model_data_parsed["model_full_name"]
# 获取对应的build_id和OBS链接
build_info = build_id_map.get(model_full_name, {})
build_id = build_info.get("build_id", "NA")
obs_url = build_info.get("obs_url", None)
# 构造前端需要的模型数据
model_for_frontend = {
"model_full_name": model_full_name,
"prebuild_id": build_id,
"obs_url": obs_url, # 新增OBS链接
"latency": model_data_parsed["om_value"], # 时延
"mean_ddr": model_data_parsed["mean_ddr"], # DDR值
"bandwidth": model_data_parsed["bandwidth"], # 带宽值
"timestamp": model_data_parsed["last_modified"]
}
performance_data["models"].append(model_for_frontend)
return performance_data
@app.route('/api/refresh', methods=['POST'])
def refresh_data_api():
"""手动触发刷新操作API"""
start_time = time.time()
try:
data = request.json
chip = data.get('chip', 'Ascend610Lite')
group = data.get('group', 'rl_nn')
logger.info(f"手动刷新请求 - 芯片: {chip}, 组: {group}")
# 1. 从.track文件夹提取build_id
build_id_map, build_missing = extract_build_ids_from_track(chip, group)
# 2. 下载所有文件
downloaded_files, download_errors = download_files_with_build_ids(build_id_map, chip, group)
# 3. 构建响应
response = {
"status": "success",
"chip": chip,
"group": group,
"downloaded_files": downloaded_files,
"download_errors": download_errors,
"build_missing": build_missing,
"process_time": round(time.time() - start_time, 4)
}
return jsonify(response)
except Exception as e:
logger.exception("刷新数据时出错")
return jsonify({
"status": "error",
"error": "服务器内部错误",
"details": str(e),
"process_time": round(time.time() - start_time, 4)
}), 500
@app.route('/api/performance', methods=['GET'])
def performance_api():
"""性能数据API接口，添加refresh参数控制下载"""
start_time = time.time()
try:
device = request.args.get('device', 'Ascend610Lite')
group = request.args.get('type', 'rl_nn')
refresh = request.args.get('refresh', 'false').lower() == 'true'
logger.info(f"性能API请求 - 设备: {device}, 组: {group}, 刷新: {refresh}")
performance_data = get_performance_data(device, group, refresh_data=refresh)
process_time = time.time() - start_time
response = jsonify({
**performance_data,
"process_time": round(process_time, 4)
})
response.headers['Cache-Control'] = 'public, max-age=300'
return response
except Exception as e:
logger.exception("处理请求时出错")
return jsonify({
"status": "error",
"error": "服务器内部错误",
"details": str(e),
"process_time": round(time.time() - start_time, 4)
}), 500
@app.route('/')
def home():
"""首页路由"""
return render_template('index.html')
if __name__ == '__main__':
# 确保下载目录存在
os.makedirs(DOWNLOAD_PATH, exist_ok=True)
# 创建所有JSON子目录
for chip, groups in JSON_PATHS.items():
for group, path in groups.items():
os.makedirs(path, exist_ok=True)
# 添加命令行参数解析
parser = argparse.ArgumentParser(description='AI芯片性能监控服务')
parser.add_argument('--host', type=str, default='127.0.0.1', help='服务监听地址')
parser.add_argument('--port', type=int, default=8080, help='服务监听端口')
parser.add_argument('--print', action='store_true', help='启动时打印所有.om和mean_ddr值到终端')
args = parser.parse_args()
# 启动时打印所有.om和mean_ddr值到终端
if args.print:
print_all_om_and_mean_ddr()
app.run(host="127.0.0.1", port=8080, debug=True)
检查后端
最新发布
扫一扫
188
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
