Six ways to prepare your core network for traffic overload
Get ready before traffic growth knocks it over.
By Bert Latamore, Computerworld | Computerworld UK | Published: 00:00, 29 August 2006
In many organisations the core network is taken for granted, largely because Cisco, the pre-eminent enterprise network vendor, has built expansion headroom into its Catalyst 6500 switches and most large organisations moved to 10Gbit/s links several years ago.
But core traffic in many organisations is growing 100 percent per year, warns Rob Whiteley, senior analyst for enterprise networking at Forrester Research. At that rate it's only a matter of time before they start running out of growth room, and in some cases may start seeing other problems with their current core technologies.
"It is much better to get ahead on potential problems than to face an emergency that requires you to take the core network down for a day," Whiteley says. Also, core network technologies are advancing at a pace matching the rest of the computer industry. For many organisations today, losing the core network for a day would be a major business emergency.
Whiteley lists the following six issues in order of importance as the key things that large organisations need to be concerned with as their network traffic doubles again this year:
1. Bandwidth is critical, but latency is king
"If you don't have a low-latency core network, then you have the wrong foundation for the rest of the kingdom," says Whiteley. Low latency is vital for switching applications efficiently and for real-time disk mirroring and other data protection and fail-over technologies.
To get low latency, large organisations have gone to semi-proprietary technologies such as Infiniband inside the data centre, with optical connections between data centres. This created a bifurcated core network at the physical level. However, in recent years the price of optical has come down dramatically, and Ciena and other industry leaders have added low-latency, high-bandwidth optical switching capability and other functionality that is making optical more attractive both inside the data centre and further out into the network.
"Now the big banks, pharmaceuticals and other large corporations are issuing RFIs [request for information] for the new optical network technologies," Whiteley says. Also, major suppliers including Cisco are now offering low-latency, high-bandwidth Ethernet that is attractive inside the data centre.
2. Intelligence at the edge means value. Intelligence at the core means complexity
The second most important issue in the core network is reliability. Core networks are designed with modularity, session fail-over and clustering, and they are built with proven, highly reliable equipment that, like Ciena, often comes out of the carrier marketplace.
However, many things, such as extra intelligence, that are important in the edge network just add problems in the core.
3. Scalability is vital as core traffic grows 100 percent a year
"Best practice is to use a platform with plenty of headroom," Whiteley says. "Cisco has done well designing the Catalyst 6500 switches to allow users to upgrade the backplane by adding more bandwidth and switch capability."
But some large organisations are reaching the limits of the Catalyst 6500. For them, he suggests looking at load-balanced switch clusters. These give 100 percent capability increases for each new box added, rather than the 50 percent that is normal without load balancing.
4. The evolution of traffic characteristics
Traditionally most traffic was unicast, point-to-point, from one user or server to one other. Video-over-IP has introduced multicast -- one origin point to multiple end-points.
A typical video conference may be from one central point to eight or more corporate locations. "What we blithely call Ego-Over-IP -- the CEO giving a morning video broadcast every day to all employees -- can involve thousands of end-points on employee desks. Without the right equipment, this could involve thousands of separate data streams over the core network," Whiteley says.
The solution is core equipment that "speaks multicast protocols" that allow the network to send one data stream and then branch it to reach multiple separate destinations as close to those users as possible. Not all core network switches and routers have multicast, however, so network administrators need to be sure they have upgraded the network before the organisation starts heavy video conferencing.
5. Managing for efficiency
Traditionally core networks were built on a static model. "You install a bunch of boxes and bring in a certified engineer or consultant to do all the black-art command-line interfacing, and the network was done. Management tools were always an afterthought of the vendors, almost a marketing effort after the hardware was built and almost ready to go to market."
Today multicasting, the modular approach to devices and the huge growth rates in traffic demand much more sophisticated network management and intervention tools such as IBM's Tivoli, HP's OpenView or the Cisco tool set. These allow the network manager to see the network topology and the growth of traffic and identify potential trouble spots. They can intervene -- adding capacity or switching to a spare router -- before trouble develops.
"Core networks are still often architected with redundant switches and routers, but this requires that traffic on any one box never exceeds 50 percent of its capacity -- in some cases 40 percent -- to provide headroom to handle the fail-overs. That becomes an expensive engineering exercise," Whiteley says.
More advanced companies see sophisticated management tools as an investment to save money in network infrastructure. Instead of designing to react to failures, they want to anticipate problems to prevent them, allowing them to design for the actual traffic loads instead of building in 100 percent over-capacity.
6. Diversity in the core network
"We get a lot of calls from companies asking how they can create a healthy amount of diversity in the core network," Whiteley says. The carriers have always split their core equipment between Cisco and Juniper Networks. Corporate strategy, however, has long been to single-source to get advantages in volume purchases and greater influence over the vendors. Cisco has grown quickly on that single-source strategy.
But in other areas, including servers and the network periphery, companies now diversify both to get greater leverage in pricing and safety against a malware or intrusion attack leveraging a single flaw in a technology, which is one reason that most companies have a mix of Linux and Windows servers. Whiteley, however, says this is a bad idea in the core network. "The cons outweigh the pros. Maintaining two skill sets, duplicate operations and the forklift upgrades needed to create that diversity make it not worth it in the core."
Finally, Whiteley is a champion of standards throughout the network, partly to allow the organisation to take advantage of new technologies easily. The issues of five years ago -- supporting the best feeds and speeds and protocols -- have been commoditised in Ethernet and other network technologies, and the competitive arena has moved upstream.
"Cisco and Juniper have moved up the stack and are providing better applications and higher level features, while Chinese companies like Huaewi are focusing on low cost/high volume production," he says. "To take advantage of these improvements as they come out, you want a standards-based network that allows you to plug-and-play."