[Galene] UDP muxing: a completely unscientific benchmark

[Galene] UDP muxing: a completely unscientific benchmark

[Juliusz Chroboczek]

Hi,

70 seconds test, one sender, 10 receivers.  The server is running on
a 4-core Cortex-A53 at 960MHz (an Allwinner A64 SoC).  I didn't think
about generating a TLS certificate in advance, so the CPU time includes
roughly 2s spent generating an RSA key.

  No UDP mux: user 20.15s, system 15.05s.
  UDP mux:    user 21.61s, system 19.05s.

I find the increase in system time suprising, but since I didn't repeat
the test, it might be an experimental error.  I've generated a CPU
profile, and the time spent in the UDP mux is negligible, on the order of
a fraction of a percent.

I think we may conclude that the overhead of UDP muxing is negligible.

-- Juliusz

[Galene] Re: UDP muxing: a completely unscientific benchmark

[Dirk-Willem van Gulik]

On 22 Mar 2025, at 20:50, Juliusz Chroboczek <jch@irif.fr> wrote:

> I think we may conclude that the overhead of UDP muxing is negligible.

Same here - I tried a standard 1 minute 256 streams x 10 tests against it - and could (on a low end VPS) not measure the difference. 

This is with Galene `behind’ an apache reverse proxy and package in a FreeBSD jail; and comparing all CPU taken by that jail. Where the latter is either configured to take it through NAT with the -udp mux option; or `normal’ with its IP directly exposed.

Dw.