On Sun, 13 Apr 2014 20:01:30 +0300, Ilari Liusvaara wrote: 
>On Sun, Apr 13, 2014 at 09:16:39AM -0400, Phillip Hallam-Baker wrote:
>...
>There's also another factor to consider: Memory use. Specifically,
>how much memory "sessions" that aren't currently being prcessed,
>but which need to be able to bring to service in 0 RTT use.
>
>I have no idea how DTLS would fare here...
>
>Also, agreed that DTLS is a serious mess (and I think it is going to get
>even worse).

A couple of comments:

- it seems to me like memory use is the biggest constraint if you're
running a server, and latency for the clients.

- unfortunately these conflict with each other: you reduce latency by
  keeping connections open and reusing them, but that increases memory
  use at servers

- We've done some studies about memory use for TCP+TLS, with real traces
  from several several different types of servers.  See
http://www.isi.edu/~johnh/PAPERS/Zhu14a.html for the tech report.
(We'd love comments and feedback!)

- the high order bit from the study is that with very conservative
  (i.e., short) timeouts, and with TCP+TLS from stub to recursive and
  TCP only from recursive to authoritative, we see ~21% greater latency
  and 9GB of state at the recursive resolver.  You can tweak the numbers
  to get different trade-offs if you prefer.

Both of these are more than current UDP, but both are, I think,
"reasonable".

And fortunately, DNS is embarrassingly parallel, so one can easily
increase server capacity by buying more servers.


About specific protocols:

- we took the most conservative choice: TCP and TLS.  Heartbleed shows
  us subtle bugs are possible, but IMHO better something the community
  has been studying for years and years than something hand-rolled,
  or something less well examined.

- Although not a privacy issue, I think DNS over TCP is a really
  interesting method to address other (non-privacy-related) security
  questions, like reducing DNS amplification attacks, and removing the
  endless debates about "change X will overflow packet size Y and cause
  subtle problem Z to W percent of users".  (IMHO, that debate has
  subtly distorted a number of policy discussions, like keysize choices.
  I'd just like to sidestep all that.)  Details are in the tech report;
  I don't want to get off-topic wrt dns-privacy.

- I think UDP+DTLS vs. TCP+TLS is an interesting open question.  It's
  not obvious to me that DTLS helps that much, as you still do a
  standard TLS handshake (just over UDP).  And I'm admiring of TCP for
  other reasons (see the prior bullet).  But I think a prudent step
  would be to advance TCP+TLS and then see if the same model also
  applies to UDP+DTLS, if we can articulate the compelling advantages of
  DTLS.


   -John Heidemann

_______________________________________________
dns-privacy mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/dns-privacy

Reply via email to