search

esp-rs / esp-wifi-sys

Wi-Fi and BT drivers packaged for integration into bare-metal esp-wifi.
Apache License 2.0
399 stars 93 forks source link

Improve response times #413

Closed karlri closed 9 months ago

karlri commented 9 months ago

When the network interface can't make any progress, it's probably a good idea to yeild the task so that the wifi task may run and proceed with actually sending/receiving data.

This seems to greatly improve response times, especially with low tick_rate_hz values. At 100Hz, ping time goes down from 25ms to <2ms. A simple http server test saw a response time reduction from 45ms to 6ms. At 1000Hz tick rate the gains are diminishing but there are some small gains. I ran my tests on esp32s3.

This might also increase throughput. I have not tested that. It also may have some unforseen adverse effects.

Ideally, we would have a more fine grained return value from interface.poll() than a bool so we have a better idea if yeilding is appropriate or not.

Perhaps with many sockets, yeilding like this is not a good idea.

MabezDev commented 9 months ago

Interesting, thanks for the PR! I wonder if this change will also improve the throughput... :thinking:. I will test it out with the bench example on Monday :).

edit: hehe I just read your description again and you already mentioned this :facepalm:

karlri commented 9 months ago

Also, I forgot to mention that it seems like the response time "relative variance" is decreased. No scientific measurements done.

bjoernQ commented 9 months ago

Nice idea! Totally makes sense

bjoernQ commented 9 months ago

I don't see such a dramatic effect but it's definitely a nice improvement at least for non-async. (in my setup the pings w/o this are 14-16ms, with the changes they are 6-8ms)

karlri commented 9 months ago

@bjoernQ What esp product do you have?

This fix does not cover all cases. Notably async and coex would probably need some work. If results are promising we can look into that aswell.

I will submit my code that i used for testing in the evening. I tested on esp32s3. My test just accepted connection on tcp, sent a short reply, flushed and closed in a loop.

bjoernQ commented 9 months ago

I tested on the original ESP32 since that one was connected to my system when I looked into it - haven't checked on any other chip yet. I tested with the dhcp example in this repo

bjoernQ commented 9 months ago

Thanks again for the great idea. I will close this in favor of #430 - I included the original commit there to make sure you contribution gets honored