[tor-talk] Why some optional weights of router bandwidths are set to 10000?

[Roger Dingledine]

For those of you wondering what the heck we're talking about: these are
parameters that clients read from the consensus that help clients choose
paths in a way that is globally optimal for load balancing.

The idea stems from the realization that relays that can handle exiting
are going to get a lot of load already by clients who choose those relays
as the final hop in their circuit, so clients should avoid picking them
in *other* hops of their circuits, because that would just overload them
even more.

More info here:
https://gitweb.torproject.org/torspec.git/tree/dir-spec.txt#n2434
https://gitweb.torproject.org/torspec.git/tree/dir-spec.txt#n3045
https://gitweb.torproject.org/torspec.git/tree/proposals/265-load-balancing-with-overhead.txt

You can see the current weights in your consensus file: search for the
"bandwidth-weights" line.

Ok. With that introduction, here we go:

On Wed, Jan 02, 2019 at 09:47:14AM +0800, Lucon Yang wrote:
> As we know, some weight params are applied to router bandwidths during path
> selection. In the process of computing node weight, I find the the
> consensus' "bwweightscale" param and the exit position related
> params(eg.Weg,Wem,Wee,Wed) are all set to '10000', which results in every
> node has the same weight value of 1. That means node's flag, such as
> 'Exit','Guard','V2dir', doesn't have effect on choosing the exit node.So I
> want to ask what's the reason of the design like this?


I admit I've always found the per-position weightings logic to be a bit
of a mystery, but I believe the answer here is that these weights are
applied *in addition to* the individual weights for the relays, and the
goal of these weights is to give clients a chance to avoid using relays
in circuit positions that aren't globally optimal for the network.

With the current relays in the network, exit capacity is scarce. So the
current values for the weights mean that if a client is considering a
given relay for whether it should be the exit relay in their circuit,
its chance of being used as the exit relay should not be reduced. Or
said another way, these weights mean that if a client is thinking of
picking that relay as its exit, it's fine to do.

Compare this situation to

         Wgd - Weight for Guard+Exit-flagged nodes in the guard Position

That weight means that if a client is thinking of picking a relay for
its Guard position, and the relay has both the Guard flag and the Exit
flag, then it should adjust its chance of picking that relay by that
factor. In this case, since Wgd is 0 in the current consensus, it should
adjust its chance down to 0, meaning that clients will skip over relays
with both the Guard flag and the Exit flag when they're picking a relay
for the guard position (first hop) of their circuit.

In an alternate world where the Tor network had plenty of spare exit
capacity, the "Weg,Wem,Wee,Wed" weights might be less than 10000,
which would instruct clients to choose these various kinds of relays
less often for the Exit position, because the relays would presumably
already be getting other load from being chosen by other clients for
non-exit positions.

Hope this explanation helped rather than making things worse :)
--Roger