September 21, 1996
By Rick Tinsley, Newbridge Networks Inc., and Larry Lang, Ipsilon Networks Inc.
ATM MPOA vs. IP Switching
Many members of the ATM Forum are betting on MPOA (Multiprotocol
over ATM) as the way to route
and bridge a full range of protocols,
including IP and IPX, over switched ATM backbones. But a growing
number of vendors are getting behind IP switching--software that
enables an ATM switch to act as an IP router. These vendors suggest
that the complexity of the MPOA spec may actually be standing in ATM's
way.
To give net managers a better sense of how these solutions
differ, Data Communications asked Rick Tinsley of Newbridge Networks,
a leading proponent of MPOA, and Larry Lang of Ipsilon Networks, an
advocate of IP switching, to take sides on the two schemes.
Larry Lang:
I think you and I would both agree, Rick,
that for ATM to gain momentum in the industry--and for it to be really
useful to network managers--it has to interoperate easily with other
technologies and protocols.
Rick Tinsley:
No argument on that score, Larry.
ATM has to play well on multiprotocol networks, which is what specs
like the ATM Forum's MPOA are designed to en
sure.
Lang:
Well, it didn't take long for us to disagree. I
would argue that convoluted software standards like MPOA are actually
holding ATM back. Most current ATM software is too complex to deliver
effective interoperable solutions. And the primary source of this
complexity is the ATM Forum, which chose to derive most of its
software standards from the ITU's Telecommunication Standardization
Sector [ITU-T]--an international consortium of telecom carriers.
Not surprisingly, given the telecom bias of the ITU-T, the forum
adopted ISDN signaling as its network access paradigm. Then, in a
compromise apparently chosen because every participating company found
it equally inconvenient, the forum decided that ATM addresses should
be obtained by converting ISDN phone numbers to OSI addresses. In
other words, all networks based on forum standards must be accessed
with ISDN signaling and OSI addresses.
So things are mighty complicated from the start. And this initial
complexity onl
y breeds more of the same. To accommodate the new ATM
addresses, the forum had to concoct a whole new routing protocol--PNNI
[private network-to-network interface]. And then familiar addressing
methodologies, like that used in IP, had to be mapped to the ATM OSI
addresses through complicated mechanisms like LAN emulation
[LANE].
All this multilayered address mapping soon becomes so thick that
ATM's QOS [quality of service] and multicast features are completely
concealed from higher protocol layers. And even with all this
sophisticated software, forum standards still require processor-based
routers to traverse subnets, a limitation that restricts ATM networks
to about 250 nodes before serious performance bottlenecks occur.
THE RIGHT ROUTE
Tinsley:
Let's be careful with those sweeping generalizations about all ATM Forum standards. Yes, LANE hides QOS from the Internet layer and requires external routers to traverse subnets. MPOA doesn't. That's the whole point. MPOA supports
routable protocols like IP and IPX and bridges all protocols at full switched speed, whether they are routable or unroutable (like Netbios and SNA). And it doesn't forgo ATM QOS and connection management features to do so. IP switching, in contrast, only routes IP. It can neither bridge nor route any other protocol.
Lang:
Actually, IP switching bypasses the forum's ISDN
signaling, OSI addresses, and all the unnecessary complexity that
specs like PNNI introduce. Instead, it puts IP software in direct
control of ATM hardware, a simplification that yields in several
advantages. Existing routing protocols like OSPF [open shortest path
first] or BGP [border gateway protocol] are used unchanged, easing
interoperability and eliminating retraining. IP addresses are mapped
directly to the underlying ATM hardware, reducing complexity and
latency. Full quality of service and multicast are made available to
new and existing applications, without massive redevelopment. Finally,
processor-based routers
no longer interfere with switched flows, thus
increasing performance, decreasing latency, and saving money.
Tinsley:
Come on, Larry, there's a serious price to pay
for IP switching's simplistic approach. In bypassing all the so-called
complexity of forum specs, IP switching also bypasses all the benefits
of connection-oriented standards like MPOA. For instance, how do you
deal with virtual LANs? And what about virtual subnets? Don't tell me
the net manager still has to configure a specific subnet mask on every
physical port.
MPOA, because it was designed with ATM in mind, enables
protocol-specific virtual LANs and facilitates moves, adds, and
changes across the whole network. Autodiscovery and autoconfiguration,
which simply aren't available with IP switching, are realities for
MPOA. In addition, since MPOA is based on standard ATM, specs like
LANE and circuit emulation can coexist on the same switched network.
But IP switching eliminates standard ATM switched virtual circuits
[SVCs
] and renders the network useless for true multiprotocol
traffic.
VIRTUAL REALITY
Lang:
But the MPOA draft spec doesn't even mention
virtual LANs, so the management you describe must be proprietary. And
even if some future draft of MPOA were to outline a virtual LAN
standard, the utility of virtual LANs for net managers remains greatly
overrated.
For example, they don't really help with adds, which still require
manual address assignments. But DHCP [dynamic host configuration
protocol] can handle changes automatically--and it comes free with
most TCP/IP implementations, Windows 95, and IP switching. What's
more, once net managers implement DHCP for adds, they also use it to
handle moves and changes. In other words, virtual LANs--another source
of needless complexity--have been rendered unnecessary before they
were ever deployed. So while it's simpler than MPOA, IP switching also
is more robust.
Further, IP switching isn't limited to IP routing in the way you
sugges
t. An IPX extension should be generally available early in 1997,
probably before the MPOA spec is finished up. As for nonroutable
protocols like SNA, the trend toward carrying this protocol within IP
is well established; IP switching simply leverages this trend. Of
course, in networks where SNA dominates IP this might be
inappropriate, but those networks probably don't require ATM anyway.
That said, a bridging extension for IP switching is under
investigation, but given the low cost of Layer 2 LAN switching,
bridging isn't a pressing concern for network designers right now.
Tinsley:
Virtual LANs are about a lot more than moves, adds,
and changes. They enable policy-based management that yields benefits
like security and QOS. What's more, MPOA also can work seamlessly with
DHCP to leverage automatic address assignment.
Lang:
Maybe, but needless lines of code and unnecessary
layers of complexity surely aren't the goal either. Besides, all of
this underlines the biggest diff
erence between IP switching and MPOA:
IP switching is a shipping reality; MPOA is an ever-receding promise.
IP switching follows the tradition of TCP/IP, emphasizing working
implementations, published protocols [RFCs 1953 and 1954], and
multivendor interoperability. A score of vendors have already
participated in IP switching interoperability tests, a milestone far
in the future for MPOA.
Tinsley:
Wow, talk about marketing hype! First, a
baseline MPOA spec exists today, which means standards-compliant
products can be built. The official goal is to go to straw ballot from
the forum's December meeting and wrap up the spec by February 1997.
Second, MPOA products have been tested and verified to deliver
wire-speed IP routing in large configurations. I challenge an IP
switch vendor to publish the same sort of performance data for its
products. [MPOA performance measurements are available at
http://www.vivid.newbridge.com.]
Third, those "informational" RFCs you mention are nice, but you
s
houldn't confuse the public by suggesting that they're ratified
industry standards. They're not even on the standards track at the
IETF [Internet Engineering Task Force]. MPOA was born of the same
process as TCP/IP--proving the concept with working product and
allowing for real-world input before finishing the spec. It even
incorporates several IETF solutions, including RFC 1483, NHRP
[next-hop routing protocol], and Mars [multicast address-resolution
service].
Finally, since MPOA is on the forum's standards track, it's
guaranteed to work over any vendor's standard ATM switches.
UP TO SPEC?
Lang:
Let's get real here. No truly MPOA-compliant
products can be produced because the spec is nowhere near completion.
Any performance numbers that have been generated for proprietary
products will undoubtedly be undermined by the whims of the MPOA
committee. Several major vendors have publicly announced support for
IP switching.
More importantly, an IP switching interoperabi
lity test was held in
August, with more than a dozen vendors participating. That sort of
progress puts IP switching far ahead of the MPOA debating
society--marketing hype notwithstanding.
Tinsley:
Let's take a look at the money. MPOA offers
unprecedented cost-effectiveness: Switched IP routing at approximately
$0.15/packet per second [pps]. Proprietary PVC [permanent virtual
circuit] schemes are considerably more expensive.
Lang:
IP switching is not a "proprietary PVC scheme." It
takes advantage of ATM's ability to set up dynamic SVCs--using routing
and addressing information native to IP.
Tinsley:
But it doesn't interoperate with ATM Forum
standards.
Lang:
Interoperability with UNI [user-network interface]
or LANE would mean setting up calls before sending the first packet.
That's contrary to the goal of IP switching. What's more, vendors
including Digital Equipment Corp. [DEC, Maynard, Mass.] have committed
to building a switch that
will simultaneously support forum signaling
and IP switching.
Further, MPOA isn't likely to offer the price-performance advantage
you claim. We've computed the cost of IP switching at just over
$0.01/pps.
GO WITH THE FLOW
Tinsley:
Creative accounting aside, Larry, the reality
is that IP switching and MPOA data-forwarding operate in a similar
manner, regardless of the size of the transaction. Both schemes
leverage a default forwarder until a flow is detected and then
initiate a cut-through circuit as appropriate. The only difference is
that the IP switch requires flow detection to occur at every hop along
the path while MPOA detects a flow once and then establishes an
end-to-end virtual circuit that may span several switches. The result
is better scalability, performance, and latency. While IP switching
may work fine over a single switch, the process of detecting a flow in
each switch along the path before setting up a cut-through connection
leads to very high latency.
I fail to understand why--once a flow is detected--the IP switched
network doesn't establish an end-to-end SVC, which is what MPOA does.
Lang:
You've missed one of the fundamental benefits of
IP switching. When flows have been identified and cut through across
multiple switches, end-to-end latency is exactly the same as that of
forum signaling. There's no need to reassemble cells into packets
between IP switches once a cut-through path is determined.
MPOA is actually at a disadvantage on multiswitch networks. To set
up a connection each switch must map the input-port VPI/VCI [virtual
path identifier/virtual circuit identifier] to the output-port
VPI/VCI. To deliver this information across the network, a packet must
pass from source to destination through every switch.
The forum invents a special packet, called a "signaling connection
request," that threads the virtual connection behind it. Meanwhile,
the initial IP packet that triggered the connection (and any
subsequ
ent IP packets) languish at the entrance to the ATM network,
waiting for the connection to be established.
IP switching, in contrast, uses the initial IP packet to thread the
connection. Eliminating that sort of setup latency may not matter much
for large file transfers, but for short transactions it's crucial.
What's more, MPOA has to translate between IP's RSVP [resource
reservation protocol] messages and ATM Forum specs. IP switches need
no translation. So IP switching can offer all the QOS features of ATM
without requiring IP applications to be rewritten.
Tinsley:
It sounds too good to be true. And it is. I'd
really like to see those performance numbers published. The truth is
that IP switching is just a shortcut for vendors without the
engineering resources to implement MPOA. It's a cut-rate technology,
and users will ultimately pay the price in lost functionality.
By replacing end-to-end ATM signaling with proprietary hop-by-hop
signaling, IP switching trades
off scalability, performance, and low
latency for "fewer lines of code." By forgoing PNNI--the routing
protocol designed specifically to leverage the QOS capabilities
inherent in ATM--IP switching sacrifices the ability to perform
QOS-intelligent routing.
MPOA, in contrast, leverages all of the benefits of ATM while
ensuring that it interoperates easily with other technologies and
protocols. If IP switching is going to give up all the benefits of
ATM, why bother converting those frames into cells?
Rick Tinsley is vice president of the Vivid business unit at Newbridge Networks Inc. (Herndon, Va.).
Larry Lang is vice president of product management at Ipsilon Networks Inc. (Palo Alto, Calif.).
[
Home
]
[
Registration
|
Subscriptions
]
[
Contact Us
|
E-Mail
]
|
