Get your protocols in the right place and your network will run
smoothly. CW360.com's Cliff Saran explains what should go
where
Early attempts at voice over IP (VoIP), "sounded like a tin can",
according to analyst Galen Schreck, in the Forrester Research paper
IP Phones: better, not cheaper, which was published in June 2001.
Without sophisticated quality of service mechanisms to guarantee
bandwidth for voice calls, he wrote, early users of VoIP services
sounded like they were "calling from the bottom of a well".
The good news is that the situation with VoIP has steadily
improved. And although making phone calls over an IP network may
not be a high priority, the techniques now used by network
equipment suppliers to improve call quality has had a knock-on
effect for applications communicating across a network. Enterprise
resource planning, databases, client/server applications, e-mail
and Web access are all vulnerable to network tantrums: improve the
network and end-users see these applications running better
too.
If a network link is congested, check the speed of the server link,
advises Charles Gallagher, director of product management at
networking giant 3Com. The easiest remedy for a slow link from the
server onto the Lan is a second network card. "If traffic is still
congested, move to Gigabit Ethernet or reconfigure the network to
reduce the effect of the bottleneck," he says.
There are two approaches to reconfiguring the network to to
maintain a high quality of service and make the best use of
available bandwidth. The first, lower level protocol, called IEEE
802.1p, operates at Layer 2 in the TCP/IP network stack. This
provides a way to split the network into virtual Lans. Priority
network traffic can then route through the faster virtual Lan while
less critical traffic can use slower ones. This technology is now
implemented, at wire speed, in most Ethernet Lan infrastructure
products. However, when network traffic is routed across a wide
area network (Wan) via an IP link then 802.1p quality of service
information is lost.
The Internet Engineering Task Force has been working on a number of
quality of service systems. The task force and industry have now
agreed that the way forward for IP quality of service is
Differentiated Services, or Diff-Serv (RFC 2474 and 2475). This
operates at the higher Layer 3 network topology, using a protocol
known as Diff-Serv codepoint to identify priority levels for
individual packets of network traffic.
Gallagher says with this technology "high priority traffic can get
processed ahead of everything else in the network". This technique
of boosting the network performance of applications is useful for
software such as SAP R/3, where end-users may be making heavy use
of the system to run month-end reports. "By giving the [SAP
traffic] a higher priority on the network, the application will get
a boost," he says.
There is, however, only so much reconfiguration a user can do,
Gallagher notes. "If a network manager runs out of bandwidth, the
network will grind to a halt." Rather than micro-manage bandwidth,
his advice is to upgrade the network.
When considering an upgrade, it is worth noting that running a
network at close to 100% utilisation is asking for trouble.
"Network utilisation will have peaks and troughs in usage. When
installing a network with a lifespan of five years, network
managers need to factor in the growth in network usage over this
period," he says.
In Gallagher's experience, this can be as much as three or four
times current network usage. "Buy as much network headroom as
possible," he advises. This may mean the network only averages 20%
utilisation today, but as Gallagher observes, "It leaves you with
80% spare capacity."
So how much headroom should you leave? In a typical 3Com customer
installation there would be a switched 10megabit per second (mbps)
Ethernet system connecting desktop PCs to the corporate Lan which
would be based on a Gigabit Ethernet backbone. Gallagher says this
should give a good deal of headroom, but when a network contains
extremely demanding applications, "some users will use Gigabit
Ethernet straight through to the desktop".
If the prospect of wiring the whole organisation with Gigabit
Ethernet does not appeal, end-user perception will prevail. When
the network is congested, applications can behave erratically and
end-users will notice. Gavin McFadyen, solution marketing manager
for IP telephony at Nortel Networks, says, "People really notice if
network packets are lost."
When such a network problem occurs, the packet needs to be re-sent.
The most visible effects of packet loss include a user's browser
software having to reload a Web page because it failed to download
correctly or broken speech during a telephone conversation using
IP.
McFadyen says, "On internal networks, packet loss should not occur.
With sufficient bandwidth end-users should experience minimal
packet loss." However, the situation changes when connecting across
a company's Wan. The problem then is that a relatively fast Lan
link, say 100mbps from a server, is squeezed down to a 2mbps link
over the Wan.
Nortel was the first supplier to provide an edge device supporting
Diff-Serv. "Many network suppliers only support the network core
which means that on the down-link between the edge of the network
and the core, packets could be lost."
For McFadyen, one of the benefits of Diff-Serv is that it will work
on existing networks. Moreover, the next generation of IP, called
IPv6, will include Diff-Serv.
With so much going for it, one wonders why users need any other
type of network quality of service system. Roger James, pre-sales
technical director at communications giant Avaya, has a theory.
"Legacy network hardware will ignore quality of service
information," he says, and this can cause problems if the latest
switching technology is installed at the network core but older
equipment is left at the network's edge.
Avaya recommends using a combination of Diff-Serv and IEEE 802.1p
to provide strong end-to-end quality of service. To use both
schemes in the IP network you need to create a mapping between
802.1p information and Diff-Serv via the switch's mapping table.
For instance when a packet comes from a PC or IP telephone via an
edge switch it may only have 802.1p priority information. The core
switch can investigate the 802.1p information, refer to its mapping
table and convert this to the equivalent Diff-Serv data. This can
then be transmitted across the Lan or Wan to provide quality of
service based on Diff-Serv.
The router and switch themselves provide some form of quality of
service. James says that in an IP-routed environment of switches
and routers, the IP routing protocols are used to find one or more
routes between two end points or network devices.
IP routing generally relies on Routing Information Protocols (Rip)
1 and 2 and Open Shortest Path First (OSPF). "Typically in a
network you will use either Rip or OSPF, not both," James explains.
So what is the difference? In his experience, most larger
installations will use OSPF as it places a lot less load on the
network infrastructure than Rip 1 and 2.
Rip uses what is known in network terminology as "a hop count" to
determine which route is the shortest between two points. In other
words, network equipment is designed such that an IP packet would
use the least number of routers and switches to reach its final
destination. James says Rip optimises the route between two paths
so it has the least hops. In addition it will keep a record of the
other "longer" routes for fail-over situations in the event the
best route has a network problem. He says Rip passes hop
information between the routers so they can build a map of hops to
locations on the network.
It is worth noting, says James, that "Rip does not take into
account the actual link speeds between hops, just the number of
hops". You can vary the impact of a hop so that the protocol can be
influenced to take a particular route.
The network mapping information is a key aspect of Rip. As network
devices continue to talk to each other they build up a "link state
map" for the area of the network they use. James says OSPF builds
this map more efficiently, since less information is passed between
routers to build the map. OSPF is also quicker to respond and
recalculate routes if a failure occurs. In most cases the routing
protocols can be left to work out routes across the network.
However, you will probably want to exert some control to ensure the
best network response by tweaking various switches and router
configurations, says James.
Now if such changes are made they apply to all IP traffic routed
across the network. "If you want to differentiate between traffic
types you have to use Diff-Serv or 802.1p at each hop," he says. A
typical set-up could prioritise network management traffic first
(routing updates to keep the network going), VoIP second and SAP
third.
James explains, "You pick the best motorway between two points [the
route] and as you get onto the motorway you have three lanes -
slow, medium and fast - to travel along." In effect, Rip and OSPF
pick the motorway; Diff-Serv and 802.1p protocols choose the lane.
"As with driving you do not want to have to change route too often
as it requires re-calculation," James says. But an easy way to
speed up or slow down is by changing lanes, ie using Diff-Serv or
802.1p to make "priority" network traffic use the fast lanes.
More often than not, good network management involves using the
right management tools. Generally, such tools are designed to keep
the network running at peak performance: faults are identified,
prioritised and fixed by their impact on the quality of the network
service.
This works very well within the confines of a corporate network.
But there is an inherent problem with IP. Predicting how traffic
will affect the quality of service on the network cannot be easily
modelled when dealing with the huge swings in traffic common on the
Internet.
Traditionally the problem for users has been that as they attempt
to drive new business through the Web they need to ensure that they
buy enough IT capacity to deal with any level of Web traffic. But
the extra IT capacity is largely under-utilised and is seen by
business leaders as a cost they cannot justify. Giga senior analyst
Will Cappelli says that as business turned towards the Internet
economy, "The need to make accurate predictions on network traffic
is more important than ever."
Earlier this year US lingerie store Victoria's Secret spent e10m
(£6.3m) on up-rating its network infrastructure to deal with peak
demands in usage for an online fashion show it was running. The
company's previous attempt bombed when the site was hit by a
massive surge in traffic. This time, the e10m investment maintained
the Web site during peak demand.
Unfortunately, as Cappelli observes, the Poisson mathematical
models traditionally used to predict network bandwidth is grossly
inaccurate at estimating how an IP network will scale. This has led
to the situation where some users are over-provisioning by buying
far more IT than they need just in case they experience a huge peak
in Internet traffic.
So Poisson is unsuitable for modelling networks based on IP. This,
says Cappelli means that enterprise management frameworks such as
Tivoli Enterprise, CA Unicenter and HP Openview are not suitable
for predicting network load on an IP network. "The major
[suppliers] are aware that traditional management software will not
be effective in the new [IP-based] world."
He says he is seeing a management architecture emerging based on
the types of capacity planning tools used by Internet and
application service providers (ISPs and ASPs). Large financial
institutions are using such tools to predict more accurately the
scalability of their IP networks.
The alternative to Poisson, Cappelli says, is using a mathematical
model based on fractals to give you a better idea of your network's
scalability but current network management tools do not support it.
Such techniques, however, are found in capacity planning software
used by ISPs and ASPs. The future may well see the emergence of a
new class of tools employing fractal techniques to model the
behaviour of IP networks and these will help network engineers
improve quality of service on the Internet sites they manage.
In the not too distant future, users will increasingly turn to VoIP
or IP telephony to provide powerful telephone-based services. To
succeed, such applications will require networks that are tightly
managed, allowing priority voice traffic to pass through without
being hampered by network congestion. Protocols such as 802.1p and
Diff-Serv allow users to prioritise network traffic today. With
these protocols and the future roll-out of IPv6, users are in a
strong position to maintain network quality of service whatever
end-users throw at them.
Glossary
802.1p - an IEEE standard for providing quality of service
within a Lan
Differentiated Services - or Diff-Serv - a way to prioritise
real time network traffic
Fractal analysis - a new way of modelling network behaviour
to take into account peak Internet traffic
Poisson - a statistical model traditionally used for
predicting network traffic
Protocol stack - a hierarchical set of network protocols.
Higher layers are aimed at passing data specific to application
software; lower layers deal with physical hardware.
Routing protocol - a technique used by routers to determine
the most efficient way to send network traffic to another device.
Examples include Routing Information Protocol and Open Shortest
Path First
TCP/IP - Transmission Control Protocol/Internet Protocol is
the networking standard for the Internet, allowing data transfer
between computer systems anywhere on the planet.
Summary
- The easiest way to improve a slow network link is to install a
second network card on the server
- Using virtual Lans a network manager can ensure critical
network traffic is given priority within the Lan environment
- When data needs to cross a Wan, Differentiated Services
provides a means to prioritise network traffic
- Network switches and routers optimise traffic flow by
calculating the "least number of hops" IP packets need to take
- It is difficult to model the behaviour of IP networks that
connect to the Internet, as peak traffic can skew estimates of
network performance.