Some applications try to increase the buffers for performance reason. The
QUIC Go implementation for instance tries to request a 2048 kiB buffer
size.
The kernel default depends on skubuf size (which is architecture
dependent), but it is memory size independet and typically around 200 kiB
(see [1]).
Other network tuning guides suggest 16MiB for 1GB ethernet, as well as
changing the default as well as maximum bufffer size (see [2]). This
conservatively increases the maximum buffer size to 4MiB.
[1]: https://elixir.bootlin.com/linux/v5.15.45/source/include/net/sock.h#L2742
[2]: https://nateware.com/2013/04/06/linux-network-tuning-for-2013/
Currently Linux has a limit of IGMP memberships of 20. When trying to
add membership to more than that, Linux fails with:
OSError: [Errno 105] No buffer space available
Allowing more memberships should not really be problematic as memory is
allocated dynamically when membership is actually added.
However, there is a protocol limit of how many memberships a host can be
in. The number of memberships needs to fit in a single group report
datagram of 64kB. In total 5459 group records fit in a datagram, but due
to IP header options this might be slightly smaller in practise.
(see https://github.com/home-assistant/core/issues/45957).
Use a limit of 1024, which should be plenty of headroom in both
directions.
Related to: https://github.com/home-assistant/core/issues/45957
