IPv6 day and tunnels

Owen DeLong owen at delong.com
Wed Jun 6 04:58:26 UTC 2012


>>> Bigger packets makes it rather circuit switching than packet
>>> switching. The way to lose.
>>> 
>>> Faster is the way to go.
>> 
>> Why only fast when you can have both big *and* fast?
> 
> Because bigger packets makes it rather circuit switching than
> packet switching, which is the way to lose.
> 

Er... No. It's attitudes like this that killed ATM.

(argument about whether the ATM cell payload should
be 64 or 128 octets lead to a mathematical compromise
decision that was completely unworkable and vastly
inferior to either choice. Unfortunately, neither the US
telcos (128) or the EU telcos (64) would give ground
and accept the other standard.)

Larger packets for sustained flows of large amounts of data
do not make it circuit switched, they make packet switching
more efficient by reducing overhead. Especially on higher
bandwidth links.

Admittedly, if you go to too large an MTU for your bps, you
can create HOL blocking issues which have the same loss
characteristics as circuit switching. However, let's say that
anything >10ms HOL blocking is our definition of bad. At
10Gbps, that's 100,000 bits or 12,500 octets. At 100Gbps,
that's 125,000 octets.

Given the combination of Moore's law and the deployment
lifecycle, designs we do today in this regard can be expected
to last ~12 years or more, so they should be prepared for
at least 16x. At 1,600 Gbps, that puts our target maximum
MTU up around 200M octets.

>> See
>> Matt's pages on raising the Internet MTU:
>> 
>> http://staff.psc.edu/mathis/MTU/
> 
> A page with too narrow perspective.
> 
>> Time on the wire is what matters,
> 
> In senses you have never imagined, yes.
> 
?

Time on the wire is what matters.

It is the primary distinction between packet and circuit switching.

>> and on a 100Gbps wire
>> you can push 6MB in 480usec. That seems more like packet
>> switching latency rather than circuit switching latency.
> 
> 100Gbps is boringly slow.
> 

His case only gets better when you go faster. Seriously, Mas,
try to keep up.

> Are you interested in only supporting slowgrams? IMHO,
> go fast or go home!
> 
> At 1Tbps optical packet switched network, there is no
> practical buffer other than fiber delay lines.

Which argues for an even larger MTU.

> If MTU is 1500B, a delay for a packet is 12ns long, delay for
> which requires 2.5m fiber. For practical packet lose probability,
> delay for tens of packets is necessary, which is not a problem.
> 
> 9000B may still be acceptable.
> 
> But, 6MB means too lengthy fiber.

6MB at 1TB/sec is 48 microseconds which is 120 km
fiber. Modern single-mode spools 10 times that size can
actually be built within reason.

> That's how time matters.
> 
> Worse, at a 10Mbps edge of a network with 1Tbps backbone,
> 6MB packets means 4.8 seconds of blocking of other packets,
> which is why it is like circuit switching.

But you wouldn't carry a 6MB MTU out to a 10 Mbps edge. You'd
drop the edge MTU to 1500. That's why Path MTU Discovery is
useful.

> Or, at a 1Tbps link in a super computer, 48usec is too much
> blocking.

In which case you would want to use a smaller MTU. However,
I doubt that anyone is likely to run a tunnel in a situation where
48microseconds is too much latency and we are talking about
tunnel MTUs here.

> 
> That's another way how time matters.
> 
> Are you interested in only supporting circuitgrams? IMHO,
> go packet or go ITU!

Sigh... I now realize that the other end of this conversation
is a human with too narrow a view.

Owen





More information about the NANOG mailing list