为何我的asyncio HTTP请求I/O等待未按预期重叠？

尝试对同一API的不同端点发起多个HTTP GET请求，预期通过asyncio + aiohttp实现I/O等待重叠，使总耗时接近最慢单个请求的耗时，但实际总耗时约等于各请求耗时之和，请求呈现串行执行的状态。使用Python 3.12.0、aiohttp==3.9.3版本，代码中无明显遗漏await语句。

代码示例：

import asyncio
import aiohttp
import time

async def fetch_data(session, url):
    """
    Fetches data from a given URL using aiohttp.
    Includes print statements to track apparent start/finish times.
    """
    start_req_time = time.perf_counter()
    print(f"[{time.time():.2f}] Starting request for: {url}")
    try:
        async with session.get(url) as response:
            response.raise_for_status() # Raise an exception for bad status codes
            data = await response.json() # Or .text(), depending on expected response
            end_req_time = time.perf_counter()
            print(f"[{time.time():.2f}] Finished request for: {url} in {end_req_time - start_req_time:.4f}s")
            return {"url": url, "status": response.status, "data_length": len(str(data))}
    except aiohttp.ClientError as e:
        print(f"[{time.time():.2f}] Error fetching {url}: {e}")
        return {"url": url, "error": str(e)}

async def main():
    # Using a list of public JSON placeholder URLs for demonstration
    # In my real app, these are different endpoints of my own API
    urls = [
        "https://jsonplaceholder.typicode.com/todos/1",
        "https://jsonplaceholder.typicode.com/todos/2",
        "https://jsonplaceholder.typicode.com/posts/1",
        "https://jsonplaceholder.typicode.com/users/1",
        "https://jsonplaceholder.typicode.com/comments/1",
    ]
    
    total_start_time = time.perf_counter()
    
    async with aiohttp.ClientSession() as session:
        tasks = []
        for url in urls:
            tasks.append(fetch_data(session, url))
        
        # This is where I expect the I/O waits of tasks to overlap
        results = await asyncio.gather(*tasks)
    
    total_end_time = time.perf_counter()
    
    print(f"\n[{time.time():.2f}] All tasks completed in {total_end_time - total_start_time:.4f} seconds.")
    print("\n--- Results ---")
    for result in results:
        print(result)

if __name__ == "__main__":
    asyncio.run(main())

运行输出模式：

[1678886400.10] Starting request for: https://jsonplaceholder.typicode.com/todos/1
[1678886400.25] Finished request for: https://jsonplaceholder.typicode.com/todos/1 in 0.1500s
[1678886400.25] Starting request for: https://jsonplaceholder.typicode.com/todos/2
[1678886400.40] Finished request for: https://jsonplaceholder.typicode.com/todos/2 in 0.1500s
[1678886400.40] Starting request for: https://jsonplaceholder.typicode.com/posts/1
[1678886400.55] Finished request for: https://jsonplaceholder.typicode.com/posts/1 in 0.1500s
... and so on ...

[1678886400.80] All tasks completed in 0.7000 seconds.

核心问题：请求串行执行，总耗时接近各请求耗时之和，未实现预期的并发I/O重叠。

• TCP连接池并发限制：aiohttp的TCPConnector默认对同一主机的并发连接数（limit_per_host）设置为2，这意味着同一域名下最多同时处理2个请求，剩余请求会排队等待连接释放，导致视觉上的串行执行。

  解决办法：创建ClientSession时显式指定TCPConnector的limit_per_host参数，提高同一主机的并发连接数：

  async with aiohttp.ClientSession(connector=aiohttp.TCPConnector(limit_per_host=10)) as session:
  

  可根据实际请求数量调整数值，确保所有请求能同时建立连接。

• DNS解析阻塞（次要可能）：如果系统默认的DNS解析是同步实现，可能会在建立连接前阻塞事件循环。可显式使用异步DNS解析器优化：

  resolver = aiohttp.AsyncResolver()
  connector = aiohttp.TCPConnector(resolver=resolver, limit_per_host=10)
  async with aiohttp.ClientSession(connector=connector) as session:
  

• 服务器端限流：如果目标API服务器本身限制了同一IP的并发请求数，即使客户端放开连接限制，服务器也会串行处理请求。这种情况需要确认API的限流规则，或调整请求速率适配规则。

修改后，所有请求会同时发起，总耗时将接近单个请求的最长耗时（如0.2秒左右）。

内容的提问来源于stack exchange，提问作者drawfine