On Tue, 27 Oct 2020 19:25:16 +0100
Tobias Waldekranz [off-list ref] wrote:

.-----. TO_CPU, FORWARD .-----. TO_CPU, FORWARD .-----.
|     +-----------------+     +-----------------+     |
| CPU |                 | sw0 |                 | sw1 |
|     +-----------------+     +-----------------+     |
'-----'    FORWARD      '-+-+-'    FORWARD      '-+-+-'
                          | |                     | |
                       swp1 swp2               swp3 swp4

So the links selected as the CPU ports will see a marginally higher load
due to all TO_CPU being sent over it. But the hashing is not that great
on this hardware anyway (DA/SA only) so some imbalance is unavoidable.

The hashing is horrible :( On Turris Omnia we have 5 user ports and 2
CPU ports, and I suspect that for most of our users there is at most
one peer MAC address on the other side of an user port. So if such a
user has 5 devices connected to each switch port, there are 5 pairs of
(DA,SA), so 2^5 = 32 different assignments of (DA,SA) pairs to CPU
ports.

With probability 2/32 = 6.25% traffic from all 5 user ports would go via
one port,
with probability 10/32 = 31.25% traffic from 4 user ports would go via
one port.

That is not good balancing :)

Marek