Re: [DNSOP] Enough latency obsession Re: Review of draft-ietf-dnsop-cookies-00

[Mukund Sivaraman <[e-mail]>]

Hi Nicholas

On Tue, Dec 16, 2014 at 02:44:40PM -0500, Nicholas Weaver wrote:
> 
> Its time to stop obsessing over latency in DNS!
> 
> DNS doesn't exist in a vacuum, but then goes to at minimum, a TCP
> handshake, and who knows what else beyond it.  Amdahl's law matters.
> 
> How many headaches would go away if all DNS is over TCP?  And how much
> would it really make a difference in Latency?

Surely a lot of problems would go away. But I don't think we can say
that latency doesn't matter. Though it is 2 roundtrips instead of 1, the
wait effectively doubles, and may increase further by a ~constant factor
during recursion. This would be conspicuous on long networks.

As DNS resolution is at the head of the batch of items that is done when
a user uses a network service, it adds to the average turnaround time of
every item on the list.

TCP performance "feels" different depending on what it is used for. On
LFNs, slow-start can throttle up fast (being a doubling throttle), and
due to the receive window TCP can deliver a lot of data quickly
vs. DNS-like UDP that restricts flow to request/response pairs.

For DNS, where there isn't a lot of data to transmit (in normal
queries), TCP connection setup is a big part of overall time to service
a request and it may not amortize well.

		Mukund