从 python 中的命名 pipe 读取时的 CPU 使用率/速度权衡

Question

I want to read from a pipe asynchronously in python. The pipe is fed continuously from another process (a javascript process). I am using pipes for a matter of speed.

I don't want to loop and see if new lines are appended to the file, because it takes 100% of the CPU. I have tried watchgod awatch , but it adds a 300ms delay. My tradeoff is an asyncio.sleep of 1ms, it takes 10% of the CPU but adds 1ms delay.

import json
import time

import aiofiles
# from watchgod import awatch


async def amain():
    async with aiofiles.open('pipe_a', mode='r') as f:
        # async for changes in awatch('pipe_a'): # CPU=0% TIME=300ms
        while True:
            async for line in f:
                x = json.loads(line)
                print(f"pong: {time.time() * 1000 - x['time']} ms")
            await asyncio.sleep(0.001) # # CPU=10% TIME=1.5ms
            # await asyncio.sleep(0) # # CPU=100% TIME=0.5ms


import asyncio

loop = asyncio.get_event_loop()

loop.run_until_complete(amain())

What is the right way to get 0.5ms delay, and 0% CPU usage?

Answer 1

Watchgod awatch is doing an async loop itself. So that's why it took 300ms as the delay of the default loop is 300ms.

On Linux, to watch file change, one must use inotify . There is an async library for inotify on python: aionotify . Waiting for file changes using notify does not block the CPU and catches the file event as soon as it appears.

The solution takes 0% CPU usage and a 0.5ms delay:

import asyncio
import aiofiles
import aionotify
import json
import time

watcher = aionotify.Watcher()
watcher.watch(alias='logs', path='pipe_a', flags=aionotify.Flags.MODIFY)

# Prepare the loop
loop = asyncio.get_event_loop()


async def amain():
    await watcher.setup(loop)
    async with aiofiles.open('pipe_a', mode='r') as f:
        try:
            while True:
                _ = await watcher.get_event()
                async for line in f:
                    x = json.loads(line)
                    print(f"pong: {time.time() * 1000 - x['time']} ms")
        finally:
            watcher.close()


loop.run_until_complete(amain())
loop.stop()
loop.close()