[tor-dev] Weaving a Faster Tor: paper/video of possible interest

David Goulet dgoulet at torproject.org
Mon Jul 19 19:24:43 UTC 2021

On 06 Jul (13:52:50), Ian Goldberg wrote:
> Hello tor-dev,
> Steve Engler (currently part of the Shadow team) and I have a paper of
> possible interest to appear at ARES 2021 next month.
> It's a design of a multi-threaded relay architecture, of possible
> particular interest when considering relay support in arti, for example
> (though the proof-of-concept implementation is based on the usual C
> codebase).
> If you're interested, here are previews of:
> Paper: https://cs.uwaterloo.ca/~iang/pubs/mttor-ares21.pdf
> Video: https://www.youtube.com/watch?v=41a6nLUJye8
> Code: https://git-crysp.uwaterloo.ca/sengler/tor-parallel-relay-conn
>       https://git-crysp.uwaterloo.ca/sengler/relay-throughput-testing


One part caught my eye in section 4.3:

  "The large amount of locking would harm the relay’s performance, and mes-
  sage passing would break some of the assumptions of Tor’s primary scheduler,
  the KIST scheduler. Rather than using a global scheduler, each local
  connection manager uses its own local scheduler which processes only the
  connections it owns."

So KIST was also put in place in order to be able to consider _all_ channels
within one scheduling loop so to properly applied EWMA scheduling that is
basically loud circuits (lots of traffic) are less prioritize from quiet ones.

Moving to a scheduler per thread (as in only handling its set of connections),
we loose that property no? And so loud circuits end up crushing quiet
circuits on the physical link?

Another thing. Looking at figure (c), it appears only "relay/edge connections"
can queue cells on a circuit half *directly* that is not using a channel. I
assume that concurrency there between a relay connection writing a cell on a
circuit half and a cell received through a channel (from another circuit half)
has been thought of? :)

I'm asking also because within Tor, there are numerous places where a cell is
enqueued on a circuit (even an OR circuit like TRUNCATE for instance) and so
those place would need to use a channel or use the way relay connections write
them (which appears to be without a channel)?

Anyhow, good read, thanks for this paper!


-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 488 bytes
Desc: not available
URL: <http://lists.torproject.org/pipermail/tor-dev/attachments/20210719/cbbedfd8/attachment.sig>

More information about the tor-dev mailing list