May 1998
By Joaanna Makris
Internet Service Providers
Guerrilla ISPs
Young guns threaten the ISP regime—but can they really give the people what they want?
|
|
|
It's a jungle out there. Ask all those ISPs locked in combat for
corporate customers. With so many potential deals (and dollars) at
stake, the only credo that counts is kill or be killed. And the
arrival of
a new breed of smaller ISPs (internet service providers)
might only intensify the fighting. These new players are young.
They're hungry. And they're ready to offer customers a way around 'Net
congestion with a range of performance-boosting enticements—from
private peering arrangements to proprietary routing to data centers.
They're also delivering services for a fraction of what big,
established players like MCI Communications Corp. (Washington, D.C.),
Sprint Corp. (Kansas City, Mo.), and Uunet Technologies Inc. (Fairfax,
Va.) charge.
Call them guerrilla ISPs. And like any rebel force, they've been
quick to win their sympathizers. Norman Estigoy, president of regional
ISP Mica.Net Ltd. (Southfield, Mich.), is one of them. As a customer
of Savvis Communications Corp. (St. Louis), he's seen big performance
improvements. "We're getting packets from Southfield to Chicago in 16
milliseconds. It takes 30 to 40 milliseconds longer with Sprint's IP
service," he says. Barry Volts, MIS manager at manufact
urer Omron
Electronics Corp. (Schaumburg, Ill.) and customer of Savvis, is
another believer. He sent out requests for proposal on a VPN
consisting of 16 remote sites connected via T1s, and "Savvis' bid was
10 times cheaper than the one from MCI," he says— $300,000 vs. $3
million.
These ISPs have also picked up some vocal critics, many of whom
wonder about the viability of providers that don't possess their own
backbones. "As an ISP, if you don't own your underlying
infrastructure, you won't be around long," says Joel Maloff, president
of the Maloff Consultancy (Dexter, Mich.). Or as John Moshier, group
manager for IP services at Sprint, puts it: "The guy driving the bus
is the guy in control."
So where does that leave corporate networkers who want to cast
their lots with the upstarts? They need to find out for themselves
whether these new players can compete. Start by assessing the three
basic techniques the new ISPs use to keep performance up: dedicated
connections to other carriers' bac
kbones through private peering and
transit peering agreements; proprietary routing for moving IP traffic
more efficiently; and data centers that can house corporate IT
resources like servers and handle such tasks as Web maintenance. Next,
check out the value-added services—like VPNs (virtual private
networks), Web hosting, remote access, IP telephony, and
videoconferencing. Finally, make sure the ISP isn't cutting corners in
charging so little: Low rates are cold comfort when the network is
down, so search out a provider that offers strong guarantees.
Rebel Forces
|
BUILD YOUR OWN CUSTOM TABLE
Table 1: A Selection of Business ISPs
|
There are now nine ISPs that fit the guerrilla description:
Abovenet Communications Inc. (San Jos
e, Calif.); Concentric Network
Corp. (Cupertino, Calif.); Digital Island Corp. (Honolulu); Exodus
Communications Inc. (Sunnyvale, Calif.); GlobaLAN Inc. (Peabody,
Mass.); Frontier Global Center Inc. (Sunnyvale); InterNAP Network
Services Corp. (Seattle); Navisite Corp. (Andover, Mass.); and Savvis.
And they have more in common than merely promising better performance
for a lower price.
Namely, they all lease their backbones from larger,
facilities-based carriers. The rented backbones run at various speeds,
from Digital Island's T1 (1.544 Mbit/s) to GlobaLAN's OC12 (622
Mbit/s). But net managers should avoid getting hung up on speed of the
backbone alone: Performance also depends on how many customers are
sharing capacity and how much traffic they're sending. And when it
comes to this information, at least, the newcomers are displaying
veteran caginess. Of those revealing numbers, InterNAP claims only 20
customers share its OC3 (155-Mbit/s) backbone. Frontier Global Center,
on the other hand, says it h
as almost a thousand customers on its
backbone, which is made up of T3 (45-Mbit/s) and OC3 links. GlobaLAN
says it has no customers on its network yet—one that is actually
isolated from traffic on the public Internet (see "Private
Matters").
Some ISPs do more than furnish the backbone. Concentric, InterNAP,
GlobaLAN, Frontier Global Center, and Savvis also resell the access
portion of the network. The rest, however, don't offer access
lines—and that can be a hassle because customers then have to buy
access to the network directly from a separate carrier.
As is the case with backbones, a variety of speeds are offered for
access lines. All ISPs reselling access lines deliver T1 speeds, while
Concentric, InterNAP, and Savvis also offer fractional T1 and T3 (45
Mbit/s). Frontier Global Center and Savvis also list frame relay
access as an option, at speeds of 56 kbit/s to 45 Mbit/s, as well as
ATM up to 155 Mbit/s. InterNAP also offers 155-Mbit/s ATM. Remote
dial-up access is available fr
om Concentric and GlobaLAN, at speeds of
14.4 to 56 kbit/s (Concentric allows its customers to dial into more
than 3,000 POPs [points of presence]; GlobaLAN, 2,500). Only GlobaLAN
offers wireless access, at speeds of up to 28.8 kbit/s.
|
Table 2: Value-Added Services
|
The ISPs that don't resell access as part of their basic
service—Abovenet, Digital Island, Exodus, and Navisite—argue that it's
become a commodity. They think focusing on value-added apps makes
better business sense—and Digital for one offers access over its
network as a value-added feature (see
Table 2
). Its Managed Bandwidth
service allows customers to manipulate bandwidth to remote locations
in 1-kbit/s increments. Customers purchase bandwidth in 16-kbit/s
increments for $1,300 month
ly and 32-kbit/s increments for $2,750
monthly. A Web interface shows the total amount of purchased
bandwidth, bandwidth distribution to each location, and the amount
available on each link.
Other providers offer a range of value-added features, from VPNs
(virtual private networks) and IP telephony to videoconferencing and
Web publishing.
Avoiding the Crowds
Regardless of the bells and whistles, though, what these new
players really stress is that they can help customers circumvent
congestion on the Internet. Where's that congestion centered? In
public NAPs (network access points)—hubbing locations where ISP and
carrier backbones are interconnected via multiple switches and
routers.
Route flapping and packet loss are the two most common NAP
problems. Route flapping arises when a router generates erroneous BGP
(Border Gateway Protocol) messages; as every new router receives a BGP
update, it resends it to other routers—creating a ripple effect. The
'Net is flooded wit
h BGP messages and traffic slows down.
As for packet loss, providers say the problem is a simple one: NAP
switches and routers have too few ports to handle the load. "In an
average day, 40 percent of packets are discarded at the NAPs," says
Barbara Ells, industry analyst at Zona Research (Redwood City,
Calif.). And packet loss leads to retransmissions—which puts the
brakes on Internet traffic. The problem is made worse by the fact that
U.S. NAPs typically handle international traffic. "Most telcos service
Asia by bringing traffic back across one cable to the U.S. and then
sending it back out to each of the countries," says Ron Higgins,
president and CEO of Digital Island. "And they sometimes oversubscribe
that one backhaul cable by as much as 20 to 1."
Peer Groups
One way some of these ISPs deal with poor NAP performance is
through private peering. Here, two providers agree to exchange traffic
at a mutually agreed upon private access point—which means they're not
restricted
to NAPs as their only places of exchange.
But meeting private peering requirements can be a challenge.
Typically, an ISP must establish at least four direct DS-3 connections
to each provider with which it peers. And smaller players say such
stringent standards are no accident; they claim large carriers set the
bar so high specifically to lock out the upstarts. "Peering is largely
a political game," says Robert Laughlin, senior technical officer at
ISP Dataxchange Network Inc. (Maclean, Va.). "The connection
requirements are merely guidelines set by the carriers to determine if
another carrier is interesting or large enough to peer with."
Numbers may help prove the point: Most guerrilla ISPs have no
private peering arrangements. Frontier Global Center has 6, while
Exodus is a big exception in that it boasts 12.
Some of the ISPs that are locked out of private peering agreements
are resorting to transit peering. Here, an ISP simply hands off all
the traffic addressed for users not on its
network to another
carrier's for delivery.
But transit peering carries a hefty price tag: about $40,000 a
month for a DS-3 to each carrier at every peering location, according
to Laughlin. So if an ISP has struck transit peering agreements with
MCI, Sprint, and Uunet, it will pay about $120,000 monthly for the
privilege.
|
Geographic Coverage
|
Right now, five of the ISPs are trying transit peering. Abovenet
has agreements with two carriers, Concentric and Savvis each have
deals with three, and Navisite has five. InterNAP has eight: It's
established T3 connections with ANS Communications Inc. (Purchase,
N.Y.), MCI, Sprint, and Uunet, and T1 links with Apex Global
Information Services (Dearborn, Mich.), GTE Internetworking
(Cambridge, Mass.), Netcom On-Line Communication Services Inc. (San
Jose,
Calif.), and PSInet Inc. (Herndon, Va.). "These peering
arrangements provide us with direct pathways to 90 percent of global
Internet destinations," says Tony Naughtin, CEO and president of
InterNAP.
Still, the number of transit agreements an ISP has says nothing
about the number of "private NAPs" it boasts. A private NAP is the
exchange point for transit peering traffic. The more private NAPs, the
more points on the 'Net where traffic can be diverted on to another
backbone. So while InterNAP may have struck a bunch of transit deals,
it has just 1 private NAP; Savvis, on the other hand, has 19.
Then again, simply avoiding the public NAP in the first place seems
to be the real benefit here. Ask Margaret McDonald, senior network
systems analyst for the Seattle Times (Seattle), who recently switched
from Northwestnet (Bellevue, Wash.) to InterNAP. "When I ran a trace
route, the latency I was used to seeing was no longer there," she
says.
But ISPs that don't use transit peering say they ha
ve a reason:
There's no control over how traffic is delivered. "There's no way to
tell whether the upstream provider is overselling capacity," says
Robert Bowman, director of backbone engineering at Exodus, which uses
only public and private peering.
And as Sprint is happy to point out, owning infrastructure confers
other benefits. "Not only do we own the physical fiber and the
routers, but we're also able to monitor, augment, and reroute
traffic," says Sprint's Moshier. "Savvis is dependent on us for
this."
Moshier may have a case. Savvis lacks a direct connection to BBN
Planet Corp. (Cambridge, Mass.), which accounts for approximately 15
percent to 20 percent of Internet subscribers. If a Savvis customer
wants access to a BBN server, traffic is routed via the backbone of a
carrier with which Savvis has transit arrangements. At this point,
Savvis has no control over how packets are routed to BBN—so the
likelihood of latency and packet loss increase.
Corporate networkers who are w
ary of public, private, and transit
peering may want to take a look at Digital Island's approach. The
ISP's network contains no NAPs at all. Instead, the provider runs a
private international network over leased connections and establishes
direct connections to ISPs in every country it has customers.
Digital Island has two data centers: one in Honolulu, which handles
traffic headed for the Asia-Pacific region and Latin America, and
another in London, which handles European traffic. From these two
nodes it leases T1 or greater lines from such providers as AT&T
(Basking Ridge, N.J.), BBN, Sprint, and Worldcom Inc. (Jackson,
Miss.).
Digital Island says this approach means lower latency. Whereas a
typical pathway across the Internet can easily involve anywhere from 5
to 15 router hops, the design of the Digital Island net helps the ISP
guarantee just 1.
Shining Paths
Avoiding NAPs is just one way these ISPs try to boost performance.
Abovenet and InterNAP, for instance, also make
enhancements to BGP-4
(border gateway protocol 4) to keep IP traffic moving, while Frontier
Global Center uses its own software to direct requests along the
fastest route between user and Web server.
The Abovenet scheme involves proprietary software called ASAP
(Asymmetric Allocation of Packets), which addresses a shortcoming of
BGP-4: its inability to "indicate the status of a router in terms of
packet loss," according to Abovenet's CTO David Rand. ASAP sends out
periodic probes to monitor the routers of providers to which the ISP's
network is linked. If packet loss of more than 8 percent is detected,
Abovenet says it can route around that carrier. The ISP temporarily
shuts down the connection and uses another backbone to carry traffic.
Still, the scheme works only in the outbound direction.
InterNAP, meanwhile, says its Assimilator routing software offers
two benefits. First, it controls the traffic route to and from a Web
site. Second, it sends traffic over the least congested route.
Assi
milator resides on private NAP servers and collects real-time
stats on packet loss, which are stored in a database for retrieval. By
monitoring the flow of packets into and out of the private NAP, the
ISP says it can determine which backbone will deliver traffic most
efficiently. Further, InterNAP claims, it can automatically modify the
path as address routes change and backbone trouble arises.
And unlike Abovenet's scheme, Assimilator also controls the flow of
inbound traffic—thanks to InterNAP's symmetrical routing scheme. Most
ISPs use asymmetrical routing; that is, they try to remove traffic
that's headed for other backbones from their own as quickly as
possible. For example, a request from a Uunet customer to attach to a
server on MCI's backbone would be passed from Uunet to MCI at the
nearest NAP, then carried over the MCI backbone for the bulk of the
trip. The return trip is the flip side: The data from the server on
MCI's network is passed to Uunet's backbone at the nearest NAP.
I
nterNAP says traversal of multiple paths over which the client and
server have no control is the problem. With symmetrical routing,
traffic runs over the same backbones in both directions. InterNAP uses
three weights to identify and direct packets. The first is an ASN
(autonomous system number), defined in routing protocols like EGP
(exterior gateway protocol), which allows ISPs to recognize the source
of any given packet. Each routing announcement indicates the number of
hops along the route to the destination IP address—so routers can send
it along the path with the fewest. The second is the BGP-4 community
tag, which marks traffic from a user's address space and dictates
which backbone will route it to its final destination. The third is a
multi-exit discriminator, which keeps traffic using the same entrance
and exit paths.
What kind of impression do such enhancements to BGP-4 make on major
providers? Just listen to Sprint's Moshier. "We could do this if we
wanted," he says. "But we don't thin
k it gives you faster
performance."
Frontier Global Center is alone among these nine ISPs in that it
implements proprietary routing to enhance Web server performance. Its
Expresslane software determines the most efficient path between user
and Web server and identifies the Web server farm best able to
distribute content. Servers examine the IP address of the user's
machine, as well as latency, hop count, and bandwidth of the pipe.
Other major carriers offer similar schemes. Genuity Inc. (San
Francisco), for instance, uses its Web Scotch and Hop Scotch services
to monitor servers and routers and direct Web requests to the server
best able to handle them.
Leave It To Them
Abovenet, Concentric, Digital Island, Exodus, Frontier Global
Center, InterNAP, and Navisite also offer Web hosting. Not only does
this let companies outsource maintenance (since their servers are
located at the data center); it also can boost reliability and quicken
response times, since content is placed cl
oser to the backbone.
Some customers say outsourcing management has saved them a bundle.
"The amount of money I would have spent on 24-hour support alone is
more than what I've paid Frontier Global Center in total in the past
two years," says Chris Christensen, vice president of engineering at
Liveworld Talk City (Campbell, Calif.), an Internet chat site.
Christensen pays about $10,000 monthly to connect 48 servers via a
10-Mbit/s connection.
But Paul Sams, director of business development for Blizzard
Entertainment (Irvine, Calif.) and an Exodus customer, says
performance was his big reason for choosing Web hosting. "We knew we
had to be as close to the Internet backbone as possible," he notes.
Exodus hosts two Blizzard Web sites.
Companies doing business internationally benefit in a different
way: They don't have to distribute and maintain multiple servers and
set contracts with ISPs across country borders. Ask Clay Shope,
operations manager at National Semiconductor Corp. (Santa C
lara,
Calif.), which colocates servers with Exodus and Digital Island. "It
would have cost us multiple tens of thousands of dollars monthly to
run international data centers ourselves," he says. Instead, he pays a
total of just $10,000 a month between Exodus and Digital Island.
Customers interested in Web-hosting services also should count the
number of data centers in the network. Abovenet operates one, while
Exodus—with eight—operates the most. It also plans to add one data
center in Japan and another in the U.K. That compares pretty well to
such major carriers as Genuity—which has six U.S. data centers and one
in London.
And when it comes to the number of data centers, more could
definitely be better. "Distributing content allows for peak traffic to
be divided over more than one location," says Zona's Ells. Exodus'
Bowman sees the issue similarly. "With fewer data centers, there's the
problem of latency over longer distances," he says. And while latency
within the U.S. is an
acceptable 65 ms, the numbers climb on
international routes. "A 300- to 400-ms delay isn't good enough for
the majority of our customers," Bowman notes.
Dollar Details
There's no doubt that networkers who choose one of these ISPs stand
to save money. The question is, how much? Getting an answer means
sorting out the pricing schemes.
|
MORE INFO
Private Matters
|
There are basically two pricing plans for Internet access: flat
rate and usage based. Concentric, InterNAP, GlobaLAN, and Savvis use
the first method, under which customers sign up for a specified amount
of bandwidth at a fixed monthly charge. GlobaLAN is the highest-priced
of the bunch, at $2,300 a month for 1.544 Mbit/s.
Usage-based billing works just as it sounds: Customers pay only fo
r
the bandwidth they consume. Concentric, InterNAP, and Savvis all offer
this option, but they calculate usage differently. Concentric charges
according to average monthly bandwidth, sampled every five minutes.
Savvis bills according to the 95th percentile; that is, it throws out
the busiest 5 percent . InterNAP has the most complicated structure:
It divides total bandwidth into tiers. A DS-3, for example, has a
first tier of 0 to 4 Mbit/s and costs $9,000 monthly; the highest tier
is 21 Mbit/s and above, and it costs $70,000 monthly. The percentage
of total traffic that falls into each tier is multiplied by its
respective rate to get the monthly total.
The downside to usage-based billing is that unless net managers
keep a close eye on things, they could find themselves with big
bills—since use is variable and hard to predict. And the rates
themselves vary widely from provider to provider. Whereas Concentric
charges $25,500 monthly for 21 Mbit/s of DS-3, Savvis charges $48,000.
Above 21 Mbit/s, t
he gap widens: While Concentric charges $40,500,
Savvis charges $55,000 and InterNAP $70,000.
But regardless of options, customers stand to save big over larger
carriers. MCI charges $2,700 monthly for T1 service, Sprint, $2,220,
and Uunet, $2,200.
A different scheme is used for Web-hosting prices, which are
divided into three components: the rack space leased on the server
farm, the connection from the server to the Internet, and the
dedicated connection from the corporate site to the server. Because
different amounts of shelf space are available, net managers need to
determine first the number of servers they want in the data center.
Monthly shelf-space prices range from $150 (Abovenet) to $600
(Frontier Global Center). Concentric and InterNAP offer a server's
worth of rack space at no cost.
As for server connections to the 'Net, prices range from $500 to
$1,000 a month for a 1-Mbit/s link, depending on whether the
connection is dedicated or shared with other servers. Monthly prices
on dedicated links between corporate HQ and the server, meanwhile,
range from $750 to $2,000 for a T1. Digital Island's prices vary
according to bandwidth and geographic location. Setup also
varies—from Abovenet's $500 to Digital Island's $2,500.
Get the Pledge
|
BUILD YOUR OWN CUSTOM TABLE
Table 3: ISP Guarantees
|
Corporate networkers may wonder if low prices raise the specter of
poor quality. The only way to make sure that's not the case? Get
guarantees (see
Table 3
).
Abovenet offers the boldest assurances of uptime: 100 percent.
Others promise uptime of higher than 99 percent—except InterNAP,
which doesn't offer guarantees at all. Exodus also has plans
to
guarantee applications availability of 98 percent to 99.6 percent,
depending on the number of data centers used.
Some ISPs also set guarantees on packet latency and loss. GlobaLAN
is the only one to offer a solid figure—150 ms or less. Exodus
and Frontier Global Center negotiate with customers on a per-case
basis, while Concentric guarantees a delay of 150 ms or less on its
leased-line traffic—but makes no such promises for the Internet
backbone. Abovenet, Digital Island, InterNAP, Navisite, and Savvis
also make no backbone guarantees.
Even so, a guarantee is only as good as what the provider offers if
it isn't met. Here, Abovenet stands out: It offers a rebate equal to
10 times the length of the outage. Savvis offers a full day's credit
for outages exceeding one hour, while Exodus offers a full day's
credit for outages of two hours. Frontier Global Center and GlobaLAN
credit customers a percentage of their monthly bill. Concentric and
Digital Island negotiate penalties on a per-
case basis.
|
CONTACT AUTHOR
jmakris@data.com
|
Net managers also need to look at network reporting. Every ISP but
GlobaLAN and Savvis furnishes this capability. Abovenet, Concentric,
and InterNAP allow net managers to measure per-port usage through a
graphical Web-based interface. Concentric shows real-time router
statistics, which allows users to view the number of simultaneous
dial-up customers and leased-line connections onto its network.
Abovenet reports the real-time status of not only the bandwidth going
into customer servers, but also its connections to each backbone
provider. That way, net managers can see for themselves they're not
getting an overloaded link.
Joanna Makris is associate editor, WAN services. She is based in New York and can be reached at
jmakris@data.com
.
|
|
|
 |
|
|
