On Tuesday, December 14, 2021, 3:39:44 AM PST, Roger Dingledine <arma@torproject.org> wrote:

On Sun, Dec 12, 2021 at 03:42:12PM +0000, Gary C. New via tor-relays wrote:
> I have a Single Tor Relay comprised of a number of Tor Nodes. I'm always interested in knowledge sharing related to Tor Loadbalancing.
> What are your thoughts on the Pros & Cons of dedicating resources to a Single, Loadbalanced Tor Relay vs Many, Unloadbalanced Tor Relays by a Tor Operator? Perhaps, a Hybrid approach?

Hi Gary,

One of the big downsides to trying to create one "meta" relay out of
many independent relays is that each independent relay will make and
maintain its own TLS connections with other relays.

So while running a fast Tor relay the normal way might end up with
roughly one connection to each other relay in the network (which could
already be many thousands of connections), doing it the way you describe
could result in many times that number of open connections. And even
if your local router and systems can handle that many open connections,
you're increasing the load on every other relay by forcing them to handle
more connections.

There will also be more bandwidth use, since each separate TLS connection
will use its own keepalive packets to try to stay connected in the
face of firewall and NAT timeouts, and also to try to foil rudimentary
netflow-based traffic analysis attacks. And because circuits are spread
out over many TLS connections, they won't get the privacy protections
from being all bundled on a single TLS connection, i.e. a network observer
will have an easier time distinguishing which circuit a given cell is for.

Which relays are you running in this way? How do you aggregate all of
the statistics/etc into a central Tor that publishes a unified relay
descriptor? I would expect your meta-relay to have some kind of bizarre
breakage, so we should check it out and see if that's true in practice.

Thanks,
--Roger