Network performance to expect?

[yanshay]

With a simple application using embassy/esp-wifi/esp-mbedtls I'm downloading (over TLS) a 2.5MB binary file from Github, and it takes around 120 seconds.

Downloading to the computer which is on the same wifi network, is close to nothing.
I see data arrives in 16Kb blocks, with around 1sec delay between such blocks. no matter what buffers I use (whether for TCP buffers or for the read operations). So if I use 4K then 4 blocks are read fast and then it waits for around a second or so.
The application does nothing except that.

Is that expected behavior and the performance I should expect?
I believe it is not a TLS issue, so asking here rather than in the esp-mbedtls repo.

I didn't try looking if there are any esp-wifi optimizations, it's just that the difference between the laptop and the esp32 is really huge that it doesn't seem like optimization could get even close.

Edit:
I also found this issue that mention a default performance similar to what I see: #1605

[bugadani]

Can you check if the solution in #3014 works for you?

[yanshay]

It doesn't have any effect.

BTW: I'm on esp32s3.

To make sure I did it properly:

    let (wifi_interface, mut controller) =
        esp_wifi::wifi::new_with_mode(init, wifi, WifiStaDevice).unwrap();

    controller.set_power_saving(PowerSaveMode::None);

[bjoernQ]

The defaults are quite conservative - see https://docs.esp-rs.org/esp-hal/esp-wifi/0.12.0/esp32s3/esp_wifi/#wifi-performance-considerations

[yanshay]

Thanks.

What's the maximum possible (that you've seen)?

Is there any set of recommended parameters? These are quite a few to check all options.

Which ones are the most influential?

[yanshay]

MAX_BURST_SIZE made the difference and is required for the other parameters to have any effect.

Eventually I've set:

ESP_WIFI_CONFIG_RX_QUEUE_SIZE="6"
ESP_WIFI_CONFIG_MAX_BURST_SIZE="12"

Is this a combination that make sense? Burst size double the queue size?

Originally I thought that the Queue would need to be at least 12 or even double to take full advantage of the burst size, but I decreased and there was no significant impact. When I reduced the burst size it did influence the performance.