Interesting chitchat on xDSL reflector re QoS/Stupid ... Doc Stafford and Jim Southworth ringing in ... lead-in below ... posted here as tutorial.
IS QUALITY OF SERVICE NECESSARY? AT&T drives to control net via technology
but AT&T Labs researcher finds simpler is better.
By David S. Isenberg
Box: [Simply adding bandwidth could turn out to be the cheapest
approach.]
AT&T carries the burdens of incumbency in a world exploding with
disruptive technology. My concept of a Stupid Network tries to
explain the disruptions that telcos must face, but AT&T still
doesn't seem to get it.
Recently Dan Sheinbein, AT&T's vice president of network
architecture & development, told the Star-Ledger (Newark, NJ)
that the Stupid Network has "not been a particularly active area
of discussion" at AT&T lately. Even more recently, Sheinbein
told me, "On balance, AT&T's network is getting smarter."
To me, the Stupid Network - the dumb transport component of
people's applications, designed simply to "deliver the bits,
stupid" - is a consequence of the new abundance created by
technology's headlong spurt. It's enabled by the Internet
Protocol (IP). (See isen.com for details.)
AT&T Chairman Mike Armstrong says he's embraced IP, but his
strategy is clearly "intelligent." He says that if AT&T
controls the interfaces, specifications, protocols, standards
and platforms of the network, it can weave them into a set of
seamless services. If AT&T could pull this off, it would be
able to hold back the rising tide of commoditization and reglue
the delaminating value proposition. But to do that, somehow
Armstrong would have to get AT&T back into the equipment game,
stamp out IP, and repeal Moore's law.
THE APPARENT NEED FOR QUALITY OF SERVICE
At the edge of Armstrong's awareness, AT&T Labs mathematician
Andrew Odlyzko is researching the economics of networks. He is
no Stupid Network ideologue. In fact, he used to believe that
the Internet needed such "intelligent" complications as Quality
of Service (QoS) and differential pricing. Both of these make
networks treat different kinds of data differently.
But now Odlyzko's research has led him to the conclusion that
simpler is better. Odlyzko, who came to Bell Labs Research 23
years ago straight from his MIT doctorate, has convinced himself
that simply adding bandwidth could "turn out to be the cheapest
approach when one considers the costs of QoS solutions for the
entire information technologies industry."
Internet telephony, introduced in 1995, made the apparent need
for QoS acute. Until then, Internet traffic consisted of email
and file transfers, and then web page information. For these
applications, fast transmission is nice, but delays do not make
them unusable. Not so with Internet telephony - people just
can't have conversations when there's more than a few hundred
milliseconds of delay.
Differential pricing is the first cousin of QoS. If you have
different levels of service, you need some motivation for people
to use the lower-grade service. Otherwise, the argument goes,
people will always use the best service whether they need to or
not.
But now Odlyzko thinks that even simple QoS schemes may be too
complex. Two years ago, he proposed a very simple QoS plan. It
used only differential pricing. He called it Paris Metro
Pricing (PMP), after the Parisian subway system of letting
people who pay more ride in "first class" cars. These cars are
physically identical, but less crowded only because they cost
more. In Odlyzko's vision, a PMP Internet would have two
identical, parallel channels, and one would be designated "first
class." It would cost more, so it'd have less traffic and
provide better service. But Odlyzko now says that administering
parallel channels would add more complexity than users or
service providers desire.
100 LANE HIGHWAY, A FEW FAST CARS
Lightly loaded networks don't need QoS. They're adequate even
for Internet telephony. Odlyzko found that on most data nets,
traffic is surprisingly light. (His analogy for the typical
corporate Intranet is "a 100-lane highway [for] a few fast
cars.") Also, he says, other work showed only 40% of Internet
congestion is due to transmission bottlenecks, and only a very
few choke points are to blame.
As intelligence migrates to the edges of the Internet, so does
network administration, Odlyzko says, "where it is wastefully
duplicated," at great expense because it requires human
expertise. He concludes that, "The complexity of the entire
Internet is so great, that the greatest imperative should be to
keep the system as simple as possible. The costs of QoS or
pricing schemes are high, and should be avoided . . . we should
seek the simplest scheme that works . . . "
And that simplest scheme, Odlyzko says, involves flat rate
pricing and over-provisioned, lightly loaded networks with a
single grade of best-effort service. This scheme takes
advantage of rapidly improving routing and transmission
technologies, and it doesn't mess with any of the properties
that made the Internet great. But it'll be a hard one for AT&T
to control.
[The article above appeared as Intelligence at the Edge #7,
which is Isenberg's monthly column in America's Network, on
March 1, 1999. Odlyzko's work can be found at
research.att.com. Isenberg (http://isen.com/)
thanks Jock Gill (http://www.penfield-gill.com) for comments on
an earlier draft. Copyright 1999 Advanstar.]
[the general rebuttal]
The "more bandwidth" versus QoS debate has been going on for a while now.
It especially heated up when gigabit Ethernet challenged ATM in the LAN a
couple of years ago. Contrary to what proponents of adding "more bandwidth"
say, continually increasing shared bandwidth to address application
performance requirements is not cost effective or easy. With the "more
bandwidth" approach, to guarantee a performance level for an application,
deterministic bandwidth allocation is the only way to ensure service
degradation (excessive packet/cell congestion and discard) does not occur
when multiple applications are transmitting data at the same time. With
deterministic bandwidth allocation, the available bandwidth must equal the
total of all the peak transmission rates of all applications.
The following DSL example is illustrative: DSLAMs cross-connect ATM
virtual circuits (VCs) terminating at customer premises to ATM switches in
a transport network. Multiple VCs are statistically multiplexed onto a
single, high-speed uplink (typically an ATM DS3 or OC-3) to an ATM switch.
In the typical case, VCs receive no QoS guarantees, and traffic is
transported using the UBR service class, which provides best effort service
only. UBR connections have no QoS or traffic contracts within a network
and are not subject to connection admission control (CAC) or usage
parameter control (UPC). The only limitation on a UBR connection is its
peak transmission rate. UBR sources can transmit cells at any rate up to
their peak rates. If there is no resource (bandwidth) available to
transport UBR cells, they are discarded. Because DSLAMs provide limited or
no QoS, oversubscription of the uplink trunk results in degraded service
for users, the level of severity depending on end user traffic patterns.
Since UBR is the most common class of service supported on DSLAM uplink
trunks, deterministic bandwidth allocation is the only way to ensure
service degradation (excessive cell congestion and discard) does not occur
when multiple users transmit data at the same time.
Deterministic bandwidth allocation does not take advantage of ATM's
inherent statistical multiplexing capabilities, and restricts utilization
of network resources. For example, an OC-3 (155 Mbps) uplink can support
155 users (VCs) at an access rate of 1 Mbps per user. Each user can
transmit data at a rate of 1 Mbps, without contending for bandwidth with
other users. If the number of users (VCs) assigned to the single OC-3
uplink is increased to 200, and each user transmits at a rate of 1 Mbps,
users will receive bandwidth on a first come first serve basis. Some users
will be unable to transmit (their cells will be discarded) until users who
cease transmitting make bandwidth available. For bursty data traffic, like
Internet browsing, deterministic bandwidth allocation is an inefficient way
to guarantee bandwidth availability for users. Large amounts of bandwidth
can sit idle until the user downloads at peak rates.
Deterministic bandwidth allocation is also costly. In the example above, in
order to guarantee 1 Mbps service to all 200 users (VCs), the service
provider would need to provision another DSLAM with an OC-3 uplink and an
additional OC-3 port on the destination ATM switch. Another DSLAM is
required because DSLAMs typically do not support more than one active DS3
or OC-3 uplink. During initial deployment of DSL access networks, where the
number of users (VCs) is small, oversubscription of DSLAM uplinks may not
cause performance degradation, depending on user traffic patterns. However,
as more and more users are brought on to the network, contention for uplink
bandwidth on DSLAMs will increase, resulting in rising cell loss ratios
(CLRs) and declining performance.
In order to take advantage of the statistical multiplexing capability of
ATM, multiple QoS parameters must be supported. Compared to increasing
network capacity by provisioning additional DSLAMs with DS3 and OC-3
uplinks in access locations and adding corresponding DS3 and OC-3 ports on
ATM switches in the transport network, implementing QoS is a cost-effective
way to maximize the amount of users supported on network resources.
With QoS, each VC can be assigned a traffic contract that specifies the
characteristics of a connection between a subscriber and the ATM network.
Under the contract each VC can be allocated an amount of bandwidth that is
less than the VC's peak transmission rate, but more than its average
transmission rate (hereinafter referred to as the guaranteed bandwidth of
the connection). As long as the sum of the guaranteed bandwidths of
connections is equal to or less than the uplink bandwidth, it is possible
for the sum of the peak transmission rates of the VCs to be greater than
the uplink bandwidth. Bandwidth efficiency derived from statistical
multiplexing increases as the guaranteed bandwidths of the VCs approach
their average transmission rates, and decreases when they approach their
peak rates. Using QoS in conjunction with statistical multiplexing enables
service providers to guarantee bandwidth availability on uplink ports to
more users (VCs) without proportionally increasing capacity, i.e., the
number of DSLAMs and DS3 and OC-3 uplink ports required to connect to
backbone ATM switches.
Applying QoS to the example above enables a single OC-3 uplink to provide
guaranteed bandwidth to more than 155 users (VCs). If the average bit rate
of 155 users is 512 kbps and their peak rate is 1 Mbps, QoS can be used to
provide guaranteed bandwidth of 768 kbps to 155 users. An additional 45
users, each receiving guaranteed bandwidth of 768 kbps, can be added to the
OC-3 uplink. All 200 users are guaranteed 768 kbps, and are able to burst
to 1 Mbps provided resources are available. Even if users one through 155
simultaneously burst to their peak rate of 1 Mbps, QoS ensures resources
are made available for users 156 through 200 to transmit at any rate up to
their guaranteed bandwidth of 768 kbps.
Tom Mitchell
Dir., Product Management
Promatory Communications |
