InformationWeek

For many companies, the answer is some form of load balancing. Load balancers are software packages that divide up incoming traffic am ong Web servers. They shunt traffic away from failed servers and let companies replace and add servers to their networks without affecting users. Load-balancing software can be installed on existing routers or servers. Companies can also buy hardware-software combinations, such as a router or server, already equipped with load-balancing software. Pricing typically starts at about $4,000.

Load balancing is increasingly important because of the unpredictability of traffic on Web sites. "You could have unbelievable traffic for three days and then nothing," says Susan Aldrich, a senior consultant at the Patricia Seybold Group in Boston. That inconsistency makes it difficult for IT managers to effectively plan for capacity. "You can't sit around and watch trends for demand," Aldrich says. "You might have to upgrade in a month or a week or a day."

Not all companies opt for Web-server load-balancing solutions, however. Some go one step further, making their back-end applications scalable s o that they run across multiple servers (see story, " Avoid The Server Bottleneck" ). This means load balancing is performed at the back end, where servers and architectures are more robust and stable, instead of at the front end, where servers are often piled on with little thought of architecture or stability. Other companies go further and design their entire systems for scalability and redundancy.

Web-server load balancing works well for First Union Corp. in Charlotte, N.C., the nation's fifth-largest bank, with assets of $202 billion. When Adi Khindaria, VP of Internet operations, began offering online banking services last April, he expected only 50 to 100 customers would sign up for Web banking daily. That turned out to be a major underestimate. Web banking caught on like wildfire, and about 500 customers a day have been signing up.

Khindaria wanted to add extra servers during peak demand without wasting CPU capacity during off-peak hours. He chose the HydraWe b Load Manager 1000 load balancer from HydraWeb Technologies Inc. in New York. He tied it in with Distributed Computing Environment middleware from the Open Group Inc. in Cambridge, Mass.

DCE has a distributed file system feature that creates a single system image of file systems, regardless of the local implementation. This makes adding new servers to the network easy when traffic gets heavy. "We just plug the new servers in, configure them as DCE clients, and we're off," Khindaria says. He can add a new server in 30 minutes; in emergencies, he's done it in less than five.

The HydraWeb LM 1000 is basically a router with load-balancing software. It aggregates servers on a network so they appear as one massive, high-speed server. It sits between the bank's Web servers and its link to the outside world. Once the additional servers are plugged in, the HydraWeb LM 1000 routes traffic equally among all servers. Between nine and 20 servers support First Union National Bank's Web site at any given time, depe nding on traffic. The bank expects to have 10% of its 13 million customers banking online by year's end, Khindaria says.

That's the kind of heavy traffic that Marc Haverland, manager of engineering and lead architect at MapQuest Publishing Group in Denver, is already familiar with. MapQuest, a business unit of GeoSystems Global Corp., provides interactive mapping technology on the Web. Visitors to MapQuest's site can get driving directions to any address in North America.

Every day, MapQuest's Web site gets 6 million to 8 million hits and generates more than 1.6 million maps on request. Traffic has been doubling every four to six months since the service went live a year ago. This has MapQuest adding a server every month. Haverland installed BIG/ip2 from F5 Labs Inc. in Seattle to handle the heavy traffic.

BIG/ip2 is a dedicated server-array controller that intelligently allocates Internet and intranet service requests across a group of content servers. The product is a hardware-software combination that sits between MapQuest's Cisco router leading to its T3 line to the Internet, and its servers. BIG/ip2 monitors each server for application availability and performance and distributes incoming traffic evenly among MapQuest's Web servers.

BIG/ip2 replaced TCP/IP's domain naming system facility. DNS has a round-robin scheme that handed the MapQuest URL over to a different server every few seconds, whether or not a server was down. All traffic went to that server until the URL was handed over again. Adding or taking servers off the network disrupted the round robin, cutting off service to Web-site visitors. Also, traffic wasn't distributed evenly among the servers, so some were overloaded while others were underutilized.

The improved load balancing afforded by BIG/ip2 has let MapQuest cut hardware spending. "We're able to make better use of the hardware that we have, and our systems administrators are much happier," Haverland says. Further, maintenance, exp ansion, and failures are completely transparent to users, he adds. MapQuest has two BIG/ip2s configured: If one goes down, traffic switches to the other within milliseconds.

'Like Jellybeans'
For some companies, ensuring that their Web sites are up is more important than keeping hardware costs down. Take Classifieds2000 Inc. in Sunnyvale, Calif. "We treat servers like jellybeans," says Peter Alley, VP of engineering. "We always make sure we have much more than we need."

This is crucial, because downtime would be fatal to Classifieds2000, which runs classified ads and banner ads on more than 120 Web sites, including those of Infoseek, Lycos, HotMail, WebCrawler, and Excite. The company uses Web Server Director Pro, a load balancer from RND Networks Inc. in Mahwah, N.J., to make sure traffic is distributed evenly among its servers.

Classifieds2000 has more than 1.5 million classified ads in its Oracle7.x database, which sits on about 30 200-MHz quad Pentium Pro servers running Windows NT . When a user clicks on a classified ad, the request is sent to Classifieds2000's servers, and information is sent back to the Web site.

Traffic to Classifieds2000's servers has been growing by nearly 40% a month for the past year. Alley monitors server workloads using Windows NT's built-in performance monitor. When workloads exceed a certain preset level, Alley adds more servers to the network and lets Web Server Director Pro take care of distributing the traffic among them. "We're massively redundant in terms of capacity, so if one of the servers should crash, our users won't even notice," Alley says.

Web Server Director Pro is a router equipped with software that handles load balancing. Alley says Classifieds2000 looked at various solutions and settled on Web Server Director Pro because it was the least expensive. The company runs two Web Server Director Pros-one online, the other offline in hot standby mode monitoring the first so it can cut in if the first one fails. "What I liked best about the product was that we bought them, we plugged them in, and we never thought about them again," Alley says. "They just work."

Built-In Redundancy
At San Francisco online stock broker Discover Brokerage Direct, redundancy is a given. "There's redundancy built into every piece of the system," says Darrell Davis, VP and director of technical operations at Discover, which does 70% of its business over the Internet. Every system and service all the way down to the network wires is duplicated, because, as Davis says, "it does you no good to have a reliable front end when your back end isn't redundant and reliable."

Discover Brokerage uses clusters of E-class Sun Microsystems servers running Solaris. While most IS departments create clusters from single servers, Davis builds clusters of clusters. The workload of any server that fails is taken over by a cluster lower down in the hierarchy, rather than just by another server. Applications are all object-oriented so the brokerage can add services and s ervers to the network as needed. Discover uses Resonate Application Dispatch from Resonate Inc. in Mountain View, Calif., to do load balancing on its secure Web servers.

Resonate Application Dispatch software schedules and prioritizes requests to the least-busy, most-available application servers behind the Web site, and lets each Web server access all application servers. It includes a Java-based interface. Davis opted for a software solution because of its mobility. "You can move it around, and adding more pieces to the system is just a mouse click," he says.

Regardless of which solution a company chooses, load balancing is critical to successful Web efforts. In today's competitive Internet environment, there's no compromising when it comes to redundancy and reliability.