wan.notes
Course notes, terrestrial WAN
Buddenberg/Aug97
Internet plumbing is made up of LANs, terrestrial (wired) WANs and radio-WANs
that are all interconnected with routers. This note addresses the terrestrial
WANs.
WANs differ from LANs in two ways:
-
- economic. WAN services are usually purchased; LANs are owned. Historically,
WAN pipes came from the telcos and WAN services from Internet service providers
(because the telcos couldn't spell packet switching correctly). The pipe
market is broadening somewhat.
-
- technical. WANs are made up of point-point pipes; LANs are shared-media.
Unicast vs broadcast. Unlike LANs, WAN protocols do not have media access
protocols because they don't need them.
The telephone company version of data communications.
Analog and voice. The analog voice history tells you where a lot
of the telco digital stuff came from and why there are some rather strange
assumptions when viewed from the point of packet switched data.
The phone company's engineering here is driven by the desire to
multiplex multiple telephone conversations onto a single wire between central
offices -- saves copper.
[insert analog_switch.gif]
The human ear can generally hear frequencies from about 30Hz to
30kHz. But passing this much frequency spread through the telephone system
was not feasible ... it turns out that if you pass the frequencies in the
range 30Hz - 3kHz, the speech is intelligible. Since this is about an order
of magnitude more efficient than full fidelity, this what POTS[1]
is built on. The voice input is trimmed to the requisite 3kHz (this is
what analog filters do), then feed it to a mixer. The mixer is also fed
with a fixed frequency from an oscillator. The mix of the two puts places
the voice information onto the fixed frequency carrier.
[insert analog_mux.gif]
If two conversations use different carrier frequencies, they can
both use the same wire for transmission to the destination central office[2].
At which point, the process is reversed and the original voice frequencies
are piped out the local loop to the end telephone set.
The analog system had the singular advantage in that it worked.
The problems all have to do with distance and scale -- solving them requires
amplifiers which amplify both signal and noise. And the more complex the
analog network gets, the more opportunities for analog filters (which are
by no means leakproof) to generate crosstalk. A digital transmission system
eliminates the worst of these problems, and the place where the payoff
exists is in the trunks between central offices. Enter the T-carriers as
digital transmission pipes ... which connect to the still-analog local
loop.
Migrating to digital pipes. The phone companies, being regulated monopolies,
upgrade their physical plants based on 1) depreciation schedules and 2)
the legacy equipment that they can't replace overnight and don't want to
obsolete lest they have to explain to public utility commissions why they've
got these white elephants around.
Digitizing physical layer - pipes with exactly two ends. Do not
confuse pipes with switches that take data out of one pipe and put it into
another. Although the mapping is crude, we're at the physical layer in
the ISO Reference Model.
-
- analog POTS (telephone line ... add a modem for data)
-
- T-carrier (e.g. T1, T3)
-
- SONET (Synchronous Optical Net) (e.g. OC3)
Conveyer belt analogy. A useful way to look at a telco digital pipe is
to consider it like a conveyer belt.
-
- buckets go by at 8000/sec (or 1 every 125usec)
-
- different capacities, multiples of 64kbits per bucket
[insert conveyer.gif]
Example:
T-1 pipes have 24 buckets per second (64kbits * 24 = 1.54Mb/s)
Where'd these numbers come from? The Nyquist sampling theorem tells
us that if you sample at twice the rate of the maximum frequency, you get
all the information. So if we sample at 60kbps, then we capture all the
sound in the old analog 3Khz. Which is why all these digital pipes come
in multiples of either 56 or 64kbps[3].
Local Loop. There are three technologies for what the phone
companies call the local loop -- the reach from the central office to your
residence or business.
ISDN pipes have 3 buckets organized as 2 B channels (64kb/s) and 1 D
channel which also has 64k (but 48k is used as overhead leaving 16k for
user applications). Consider ISDN as moving the analog-to-digital conversion
from the central office out to your handset. Now a 3-bucket conveyer belt
delivers 64kbit sized buckets of data to the central office where the contents
are simply shifted into the same-sized trunk buckets to the next central
office.
All carriers, (T-carriers, ISDN, O-Carriers) run at 8kbuckets/second
-- driven by the uncompressed voice assumption -- and have bucket sizes
that are multiples of 64kbit. This allows 'channels' to be multiplexed
into larger trunk lines. For instance, the two B channels of a residential
ISDN connection (BRI, 2B+D) can be mapped onto two of the 24 T-1 channels.
(Hint: PRI = 23B+D)
The big attractiveness to the telcos is that T-carriers (and later OC-carriers)
can be swapped in for analog trunk lines on a wholly incremental basis.
The shift from analog to digital does not change the circuit switch paradigm
one whit.
Other WAN technologies. Pipes from other than the
telcos.
Communications to the home has heretofore been a monopoly of the telephone
companies (and the post office). But that is -- at least technologically
-- changing.
xDSL. Digital Subscriber Line technology is an alternative to ISDN for
local communications. The way it works harks back to the POTS description
above. The local loop telephone line is frequency division multiplexed
with several 3kHz channels, much the same way analog trunk lines used to
be. Each of these analog channels can then be used to haul bits. Commonly,
the first one (the 0-3kHz one) is used to carry voice telephone calls --
POTS reborn. The rest have digital signal processors added to each end
to provide data transmission and recovery service and a multiplexer that
aggregates all these individual sub-pipes together.
By varying the number of send and receive pipes, the number of bits
carried in each direction can be toyed with. This is where the variations
of x in xDSL came from. ADSL stands for Asymmetric -- an observation that
a WWW surfer from home is likely to need only one of these channels to
pass URLs upstream; the rest are available to pass pictures of Mars downstream
to a browser.
There are distance limitations to xDSL and the limit is soft --
the bit rate supported decreases as the distance increases. But the advertising
hype indicates you should get some reasonable service for runs of around
a mile at capacities equivalent to T1.
Telephone empire crumbling. The initial business model was that
users would have an xDSL 'modem' at home, use telco local loop to the central
office switch, whereupon the bits would join the telco's cable structure.
And xDSL service offerings would appear on the telcos rollout schedule.
[insert xDSL_telco.gif]
But the barriers to entry are low and there's a buck to be made ...
an alternative is appearing: Get a 'burglar alarm' circuit that is a hard-wired
(telcos call this a 'dry splice') pair that runs between your home modem
and your neighborhood ISP's router. All the phone company is providing
is the copper. You then buy the xDSL end equipment from an electronics
manufacturer out there and connect your home computer to the ISP's router.
[insert xDSL_bypass.gif]
Note that the standardization requirement is quite limited --
the xDSL equipment must operate correctly at both ends of the wire and
the xDSL equipment at the ISP must interfaced with the router. But you
don't need world-wide standardization (like ISDN did) to get worldwide
access and interoperability over the Internet. Note that the telco's central
office switch is simply not part of this infrastructure at all. Not surprisingly,
at least some telcos are starting to make the burglar alarm circuits hard
to get.
Cable TV. Two thirds of the homes in the United States are visited
by the CATV cable plant. The fact that the cable plant is there is fairly
well known. And the cable plant itself has both
-
- capacity --capable of carrying large chunks of data.
-
- interactivity.
This by no means solves all the problems, either technically or in terms
of business model[4].
There are some interesting electrical engineering problems with internet
over cable tv, but most are known and none appear to be showstoppers in
the vast majority of cases. A few:
-
- one-way amplifiers. Most of the CATV infrastructure has a fanout splitters
and transformers that pass signals only in the downstream direction. Replacing
all these is a significant effort (although far less than restringing the
cable). At least one company offers a finesse by mating phone line (upstream)
and CATV (downstream).
-
- significant signal processing. Much the same magic that makes xDSL work
is required here. Some of the analog components in the cable plant induce
frequency drifts with temperature changes.
-
- the CATV head end must connect to an ISP. And often have a server capacity
of its own.
-
- network management requires dynamic assignment of IP addresses to the
cable modems and SNMP management to be manageable at all.
Technically, the Internet over CATV plumbing is looking very much like
ethernet today. Which means that the channel between you and the Internet
will not be yours alone -- you share it with your neighbors. But the whole-subnet
capacity will be in the same neighborhood as 10Mbps ethernet. The tinkertoy
assembly is also straightforward -- you'll probably get a box with an RJ45
UTP plug on the side that accepts and produces ethernet frames ... which
you can plug into either your computer or a hub.
Note also that a CATV 'LAN' doesn't need (indeed, can't use) a switch;
its head-end plugs directly into a router. xDSL is similar, although there's
a service offering fight in the industry at the moment: an xDSL circuit
can
go through a telco switch, but it's really an end system-to-router
hardwire link so there's no value added by the switch unless you as a user
intend to dial in to multiple routers.
Many of the problems here are rooted in the fact that the CATV
industry is wedded to the television industry -- one of the more notorious
stovepipes. Most cable companies don't know the difference between internet
connectivity and internet services.
-
- Domain name service is absolutely required over the CATV plant providing
internet services. Indeed, the assemblies that I've seen use 'cable modems'
that are very analogous to intelligent hubs -- they have an IP address
for management reasons, which is obtained using BOOTP.
-
- we'll probably get flat-rate internet access from the CATV folks because
that's the only way they can bill.
-
- the industry got hung up on ATM for a while, which impeded real progress.
-
- the industry also hypnotized itself into believing that they needed to
string fiber to every home in order to make the system work.
-
- regulation and municipal franchises often have the perverse effect of
stifling innovation[5].
Wannabes -- power line carriers. Every two or three years somebody
comes up with the bright idea that we can use the power lines that enter
homes and businesses to carry data; not a new idea. And several companies
have started up and died trying to exploit this market ... never quite
seems to take off.
Why this is a good idea. The wiring already exists in and to your
house -- you can hook up a LAN by simply inserting a gadget between your
computer's NIC and the 110V receptacle on the wall. Indeed, with a bit
of integration, the NIC could be attached to the power cable on your computer
so when you plug in the volts, you also plug into the LAN. Technically
this is quite feasible for small, low capacity, fairly simple networks.
The practical problem has to do with household wiring and all the other
things that hang off it:
-
- when the refrigerator kicks in, it puts out a sizable spike on the power
line. Which is why computers have surge protectors on the power lines.
And why UPSs have power conditioning in them, not just battery backup.
-
- network cabling is designed to eliminate splices as much as feasible.
And when there are splices they are done very precisely in order to minimize
reflections in the signal. By contrast, household wiring has lots of splices
and wirenuts. What may look like a solid connection to a DC or 60Hz AC
voltage may have pretty high resistance to a higher frequency (i.e. LAN)
signal. Further, household power wiring is not twisted so the electromagnetic
cancelling that's carefully designed into the Cat 5 LAN cabling standard
doesn't exist.
-
- parallels to what's going on inside the house are also in the neighborhood.
Generally, the networking signals will not transit the step-down transformers
on the power pole on the street.
None of these difficulties are insoluble, but the cost to do so has not
been less than alternatives.
Data link layer -- switches -- and the backbone
WAN.
Getting through the WAN cloud from your building's router to a destination
router is the subject here. How do we connect these various links so they
accept and produce datagrams to the appropriate router?
There are about four ways to build the WAN component of an internet.
Note that they're not mutually exclusive: all these technologies have router
interfaces.
-
IP over a switched network. ATM, Frame Relay, X.25 (obsolete), SMDS
(also ran) networks all fall in this category. In practice most internet
setups are in the form of Permanent Virtual Circuits through the switching
fabric so there is little value added by the switching ... but that's all
you could get from the telco. PVCs carve out a pipe for your exclusive
use through the trunks and switches and the pipe is the same size all the
way through.
-
IP over SONET. There is no switching layer between the routers and
the SONET multiplexing. Last year's high capacity routers operated
in this fashion in the OC-3 through OC-24 capacity area. The reason
for the multiplexing is that the fiber is carrying somebody else's circuit
as well as yours. Note that multiplexing is different than switching.
-
IP over 'glass'. At high capacities, this tends to mean IP-over-DWDM
(Dense Wavelength Division Multiplexing). None of the SONET multiplexing
is used, only the SONET framing. Other than in-line repeaters (for
distance regeneration) there is nothing between a pair of routers that
have DWDM interface cards in them. At the time of this writing, there
are DWDM-capable routers through OC-48 capacity. Note that the router-to-router
link is the sole user of this particular fiber.
IP over DWDM is becoming popular with large ISPs for several reasons, bypassing
the telco's old technology being only one of them. A practical reason
is that the terrestrial WAN becomes simpler: there are fewer boxes with
transistors in them between routers. And since routers -- especially
the high end ones that support DWDM -- all have SNMP agents in them, detecting
and isolating faults becomes fairly straightforward. This "Roll-your-own
WAN" simply interconnects routers with nail-up pipes (no switching at all).
Typically, these routers use SLIP or PPP as the router-to-router protocol[6]
.
With those caveats aside, lets look at some WAN switching. Unless
you are a sizable ISP, you probably won't be in the IP-over-SONET or IP-over-glass
markets.
Telco-provided switching comes in about four types:
-
- X.25 Set of X.25 switches inside the service-provider-owned switching
cloud.
-
- Frame Relay Organized same as X.25
-
- SMDS (Switched Multimegabit Data Service)
-
- ATM. Organized same as X.25
The quick hit: Frame Relay is a well understood, well documented and stable
technology. It's widely offered by both the RBOCs and several other service
providers. Because it's ideal for a router-connected internet, its popular
with users. If in doubt, Frame Relay is a very low risk place to start
your WAN.
Historical thread. Like the evolution of transmission pipes,
the history of switch evolution is useful too.
-
A. Pre-ARPANet. X.25 protocol has been used for years as interchange between
dumb terminals and mainframes. Note this is a simple network, no concatenation
of multiple nets, no routers yet.
-
B. In early ARPANet, Mil-1822 protocol was fashioned. Somewhat more efficient
than X.25 but like rest of ARPANet in those days, it was invented in-house.
No 'vendors' in the conventional sense. (Implementation is simply some
software in a general purpose computer. PDP-11s and VAXen were common victims.
No special purpose hardware is required).
-
C. When DCA (now DISA) took over the operational portion of ARPANet (and
renamed this segment DDN), it decreed use of X.25 as WAN protocol. Only
vestiges of 1822 remain in use. At this stage, the backbone in both the
research ARPANET and the operational DDN are 64kbps capacity. T1 backbones
are still a dream.
-
D. [ insert 5 years of improvement in 'pipe' technology here -- 1985-1990.
The Internet ramps up from 56k backbones to T-1 and higher]
-
E. X.25 works over T-1 sized pipes, but is very inefficient. Reasons:
-
1. Number of packets 'in flight' increases with larger pipes [7]
-
2. X.25, because of it's pre-TCP/IP origin, duplicates a lot of functionality
of TCP and is connection oriented -- but only within the X.25 domain, not
across the routers to end systems. This duplication of functionality inhibits
scale-up. These are protocol shortcomings ... can't be beat into submission
with CPU horsepower.
-
F. Frame Relay re-emerges. Frame relay, as a protocol, is really Mil-1822
risen like a phoenix from the ashes. Frame relay strips off all the connection-oriented
functions and the standard eliminated any functionality that is provided
by other layers (principally transport). It is a queue-and-forward affair.
-
1. Permanent Virtual Circuits are typically identified between routers
and programmed into the FR switch configurations. This has the effect of
making the frame relay cloud look like the roll-your-own router-router
configuration noted above. While the value added by the Frame Relay switch
is questionable, this saves you the hassle of having to educate the telco
about your network.
-
2. Switched virtual circuits are duly acknowledged in the standards, but
are not widely offered by the telcos (PacBell does not offer SVC Frame
Relay).
-
G. Also rans. IEEE 802.6 (aka Dual Queue Dual Bus) and SMDS (Switched Multimegabit
Data Service). Technically sound, but little market penetration.
Vendor offerings ... not the same as the standards.
Pacific Bell has Frame Relay switches at various central offices. Their
service offering looks like this:
-
56k Frame Relay = nailup 56k attached to FR switch
-
128k Frame Relay (or better, up through 1.54M capacity) =
T-1 physical pipe attached to FR switch. Typically you pay for a minimum
pipe size (e.g. 128kpbs) and the first 128kbps of offered traffic is tagged
with Discard Ineligible flags -- FR switches are not allowed to throw them
away when congested. The remainder of the T1 pipe -- the rest of the buckets
on the conveyer belt -- is available to you, but any traffic that uses
them has the Discard Eligible flag set; switches can throw those frames
on the deck when congested. This, of course, is not as big a deal as the
telcos think it is because TCP detects and accounts for missing datagrams.
These Discard Eligible and Ineligible flags are called Committed Information
Rate; PacBell supports CIR; many other RBOCs don't.
AT&T and MCI support Frame Relay over ISDN; PacBell doesn't.
Last year's Next Big Thing (tm).
Unmet WAN requirements with existing protocols include:
-
- 'bandwidth on demand'. (Warning: lots of definitions)
-
- Holy Grail of audio, video, data in same pipe, (the MBONE work within
the conventional IP internet aims at the same goal)
-
- more bits (a phony)
-
- limitations of existing Internet switching:
-
1) switching is done by customers, not the phone company ... so phone company
doesn't get the revenue stream.
-
2) conventional routing, and variable packet sizes requires RAM transfers.
Practical limitations on switching speeds.
I'd suggest that there are several other requirements, mostly revolving
around multicasting, that are also not met. But these are not yet widely
recognized (and certainly weren't in the ATM hype). So, we'll live with
the above list for this discussion.
There's a quite plausible story that illustrates the intellectual
origin of ATM: The switch folks at Bellcore were watching the transmission
people figure out how to multiplex more and more data into the OC pipes.
They realized that none of their switching technologies could handle the
drinks from those firehoses and needed a new approach.
Ergo, ATM.
-
- fixed cell size allows entirely hardware-based switching using registers
organized as Banyan[8]
switches.
-
- deterministic delivery. Identically addressed cells travel same route
through switching fabric. (This is only true in absence of congestion that
causes cell collisions).
Necessary dehype.
-
- cell size is a compromise for non-technical reasons that makes nobody
happy,
-
- the switching solves only a fraction of the problems. ATM is a data link
layer protocol (only). And it doesn't even cover all of the data link layer.
-
- 'bigger pipes' problems can be solved without new switching technology.
The 48 byte ATM cell payload isn't a power of two; rather it's a sawing
up of the baby -- a compromise between 64 bytes and 32 bytes. Meanwhile
'gigabit routers' have become a practical reality in the Internet core.
If we can do that, why do we need all these messy switches out in the WAN
cloud anyway?
D. I used to think that ATM was probably doomed to success due bandwagon
effect[9].
Technical shortcomings beaten into submission with brute force:
-
- Ma Bell (SONET + ATM = B-ISDN). The RBOCs are investing in ATM, but primarily
for their core switching functions.
-
- every router vendor. In the early 90s every router vendor who planned
to stay in business figured he'd better have an ATM interface. So they
all announced; real product availability took a while longer.
-
- Cable TV, Grand Alliance. CableLabs, the R&D organization for the
CATV industry, was touting ATM over the coax as the wave of the digital
television future. This has rather worn off; CATV 'cable modems'
are organized as LANs and look a whole lot like ethernet.
-
- IBM and dozen associates. 25Mb/s LAN vision. Reinvented the ATM LANs
as an upgrade path for IEEE 802.5 token rings. This has since appeared
to fall flat in the industry.
-
- Teledesic. This satellite constellation was planning to use ATM switching
for the inter-satellite links.
E. Architectural note: ATM represents a different paradigm. Cell switching
requires homogeneaity at the data link layer.
-
- early WAN implementations will be classical router-based ones. Heterogeneous
internets. This means that ATM becomes an adjoining data link layer protocol
just like LAN protocols.
-
- to get the benefits of ATM, it must go all the way to the end system
-- routers are not in the picture. This requires a lot of infrastructure
rework. Nobody believes this will happen any more.
-
ATM has found one useful niche: as a multiplexing scheme. Pacific
Bell's DSL offering (whether bought directly from the telco or through
an ISP) runs DSL from your residence to the central office at which point
the DSL modem is terminated into an ATM switch to get your datagrams the
rest of the way to the ISP's router.
Three Dirty Little Secrets about ATM ATM, while appropriate technology
for the applications it was originally designed for, has three severe architectural
shortcomings: in availability, multicasting and media access. All three
of these requirements are critical to military applications.
Availability and survivability. One fundamental characteristic of router-based
internetworks is that broken routers or links can be routed around (assuming
rich enough connectivity). If a router fails, its adjacent routers will
simply not use the links to tht router, and the end-system TCP implementations
will service for any packets that were lost with the failed router. Similarly,
in local area networks, an FDDI DAS LAN will wrap on failure of a node
reconstituting the network. All this is transparent to applications. This
is a primary characteristic of connectionless IP networks and TCP state
machines implemented on participating end systems.
ATM, conversely, is a connection oriented protocol. All cells
associated with a particular session take the same route through the switching
fabric. This means that a switch failure along this route during a session
results in a failed session. This is anything but transparent to the application
-- it must re-establish the session.
Why is this important in military systems? The designers of the
ARPANET consciously created a basic design that is highly survivable, and
automatically uses all the redundant connectivity available to it. They
were right.
Switch vendors say that ATM SVCs will transparently re-establish,
but PVCs definitely will not.
Multicasting. In real world networks, a sensor always sends data to
more than one decision support node and decision support nodes are fed
by multiple sensors -- a many-to-many relationship. You can easily support
this with two arguments:
-
- survivability. Nodes are replicated in case you lose one.
-
- common tactical picture. You want the same data in multiple places because
multiple people are looking at it.
ATM simply doesn't multicast. In theory, the standard is being patched
to fix this. On the other hand, LAN protocols inherently multicast (an
ethernet frame dumped onto the LAN is 'seen' by all the other nodes on
that LAN). Routers running Multicast IP also multicast by replicating datagrams
as required to efficiently reach all necessary parties.
Media access. This shortcoming becomes evident in radio-WANs where the
shared media requirements attend an implementation of any scale. Because
multiple radios share the same media, a means of disciplining them is required
so they don't all transmit at once. ATM, like all WAN protocols, lacks
such a media access protocol entirely.
Telephone company networks consist of point-to-point links connecting
switches that switch calls, cells, or frames. Consequently, there is no
need for media access protocols, and indeed, they do not exist in POTS,
X.25, Frame Relay ... or ATM[10].
While there are no particular objections to ATM switching in terrestrial
WANs, DoD programs should be wary of transposing ATM terrestrial technology
into radio-WANs. The 'prove it' checklist should definitely include these
points.
Precise language.
As a final thought, we may use some precision of language to sort out discussions
of pipes and switches.
Speed. It is appropriate to use the term 'speed' in guaging the
capability of a swtich (or router or hub). As in CPU speed, number of instructions
processed per second, number of packets processed per second, ... It is
not appropriate to use the term with reference to transmission.
Capacity. Transmission paths, whether copper, fiber, or ether,
are properly guaged using the term capacity. A T1 pipe has a capacity of
1.5Mbps; an FDDI LAN has a physical layer capacity of 100Mbps, etc. But
the bits go down the wire at exactly the same speed (in copper and fiber,
about 2/3 the speed of light) in both of these pipes. Indeed, so do the
bits travelling down the wire from your 14.4 kbps modem. Even an OC24 pipe
with a 4Gbps capacity still has the bits travelling at about the same speed.
The difference is that in the OC24 pipe, there are a whole lot more of
them multiplexed into the same conveyer belt.
[1]Lingo
for Plain Old Telephone Service
[2]Analog
radio works the same way, except there's a second reason to move the frequency
up. In addition to placing the information onto it's own frequency, the
carrier mixing produces a frequency that can actually be transmitted over
the ether -- audio frequencies are too low.
[3]There's
some arcane bitstuffing explanations behind the differences between 56k
and 64k -- four our purposes, this is rounding error. And all of these
sizing decisions were locked into concrete before any voice compression
technology leaked into the telcos' life.
[4]One
cynical observation I've heard is that CATV technicians were the ones too
dumb to get telco jobs.
[5]Although
the reverse seems to be true in the local dealings between the City of
Monterey and TCI
[6]Router-router
interconnect was the original motivation for SLIP and later PPP. The dial-in
from home use for PPP came later.
[7]See
Malamud, Stacks: Interoperability in Today's Computer Networks Prentice-Hall:
1992 for good discussion. This problem is a critical protocol problem in
radio-based networks too.
[8]The
tree, not the company that adopted the name.
[9]This
comment made it into my notes before the hype wore off. ATM was touted
as technology that would cure baldness for the first half of the 90s. Reality
checked in about 1995 and ATM is now recognized as a core WAN switching
technology -- what it was originally conceived for -- and of limited use
beyond that. Naturally, the word takes time to get out ... parts of the
DoD were really seduced by the hype and haven't yet recovered.
[10]It
is fair to note that this is a shortcoming of all wide area network protocols,
not just ATM.