[nycbug-talk] nyetwork neutrality, rehashed
carton at Ivy.NET
Sat Nov 1 23:57:39 EDT 2008
>>>>> "ap" == Alex Pilosov <alex at pilosoft.com> writes:
QoS is only needed on the ends because the core network by convention
does not fill. There is not and probably never needs to be
``end-to-end QoS.'' For those cases when it does fill, it's within
user expectation to have problems, like the old days when long
distance carriers reported ``there are no lines available'' after a
major televised earthquake made too many concerned relatives place
calls at once.
Corporate WAN's built from private leased lines _do_ make end-to-end
QoS a reality, today and routinely, but the technology faces a
different task for that job than for the public Internet because a
corporate WAN is a nationwide core network that has one-to-few
last-miles spoking off each POP, so it's plausible and cost-effective
to have core links that may fill before the last-mile links, while
when this happens to an ISP it's a problem that should be fixed, and
not by QoS.
I would like (and, in a rickety way, have rigged up for myself) a type
of lastmile QoS that doesn't exist on these corporate WANs and would
be hard to implement scalably, in ASIC, or across trust boundaries.
* let web browsing remain fast even while I'm bittorrenting
* let me share Internet with neighbors without risking making my own
* divide bandwidth among the people inside my household evenly
regardless of what they're doing
The type of QoS I imagine neutrality legislating would be much simpler
and less capable than this. It'd be DiffServ-ish and based on CIR.
Traffic can be Red, Green, or Black. There are three queues in the
ASIC, and the ASIC always sends Green traffic (``CIR'') if it has
some, then Red traffic (``best-effort''), and if there is no Green or
Red traffic then it can send Black traffic (``violation''). If the
ASIC has only two queues, it may drop Black traffic unconditionally
without fucking up the overall system. On Red traffic the ASIC does
RED (Random Early Detection, statistical packet dropping rather than
FIFO to make TCP less bursty). It'd be wonderful if it could do
WFQ-RED, but i don't think that will be possible with line-rate
ASIC's and is definitely impossible with L2 ASIC's.
Each packet is marked Red or Green or Black based on two checks:
1. an L3 classifier. Customer specify a list of them. If a packet
matches, it gets marked Green. Everything that doesn't match a
classifier rule is Red.
2. a rate-policer. If packets match the L3 rule, but are arriving
more quickly than the rate you specify, they get marked Black.
This scheme is simple enough to work with Unsolicited Grant. Also the
classifier and rate-policer can run on a separate box from the ASIC
having 2 or 3 queues, so the ASIC that actually implements the
priorities and does the stochastic dropping can be an L2 ASIC rather
than L3, can be managed by a different company, does not need to speak
reservation-protocol to the customer.
VoIP will of course be Green. You'll install rules on both ends of
your last mile link when you set up the call legs.
When you subscribe to receive a multicast flow (a television channel),
you'll always add a classifier to turn it Green because multicast
traffic can't be congestion-controlled like TCP (there is no closed
loop) and often won't be retransmitted at all. The Black category
gives you enough power to prevent subscribing to an unexpectedly
high-traffic multicast stream from overwhelming your pipe.
It doesn't do the link-sharing things I dreamed of to the Red traffic,
but it's enough to give non-ISP VoIP and IPTV companies
competitively-fair access to customers of Speakeasy, Time Warner,
Comcast, FiOS that are already using the kind of QoS I've just
described to provide their ``tripple play''. Since this type of QoS
is already _working_, there's an existence proof that ADSLAM's and
6500's and broadcast L1's that these four companies have implemented
on a national scale. What's possibly missing is a reservation system
for installing the filters that can work across a customer/ISP trust
boundary, and desktop free software capable to order the reservations.
Maybe there does exist such a protocol which is adequate (RSVP?), and
I just don't understand it yet.
Here is one way to implement the reservation protocol. I hope it's
not the way that Obama's team forces into legislative reality, but
describing it proves that such a scheme is within reach.
The Upstream router looks at the DSCP the customer has put in his
packets. It says either Red or Green. The upstream router does no
policing, no step 2, and expects the customer to do this himself on
the desktop. (not good enough IMHO, but a start.)
The Downstream router does flow-cacheing on packets received from the
customer. Each cached flow also measures the rate of traffic received
from the customer. The DSCP value the customer sent is cached along
with the flow in the flow-forwarding hash table. The DSCP marked in
Downstream customer-bound packets is ignored, and they're treated
according to the DSCP stored in the flow cache. Downstream Green
packets are policed to the measured rate of upstream packets in the
It's primitive. It's not adequate for IPTV. I think an explicit UDP
reservation protocol is better. but I described something well within
reach of current ASICs, sufficient for VoIP, requiring no nonexistent
desktop software, and basically abuse-proof.
Whether something's a monopoly or not has become complicated, not a
simple question of ``do they have competitors? more than one
competitor? active shoppers?'' Talk to the same people espousing
this simple view about their own plans for expansion and success in
business, and they use words like ``horizontal and vertical
integration,'' ``leverage,'' ``cannibalize.'' They're tying products
together deliberately to smother customer choices and throw sand into
the market's gears. This isn't an accident or a byproduct, but rather
their primary intent and method---their success is, _to their own view
Mr. Tripple Play,_ related primarily to how _much_ sand they can dream
into existence and how sticky it makes the gears. Market idealism is
most thorougly obsolete among the people who love, study, and succeed
in markets. Causing markets to behave less ideally in your favour is
what studying business is _about_ right now.
It's too late to treat regulation as a response to an abberation.
Maybe it used to be that, but not any more. Markets everywhere are
saturated with regulation, and are not necessarily always impeded by
it---the sophisticated markets all _depend_ on complicated regulation
to function smoothly. This is the meaning of ``post-capitalist,'' and
if there's one biggest-lesson from the present economic crisis, what
the fuck else is it?!
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Size: 304 bytes
Desc: not available
More information about the talk