Scalable architecture aids Weather.com's site
Surges in web site traffic come with the e-business territory. But few businesses have quite as much experience with unpredictable traffic floods as Weather.com.
The Atlanta-based site, which is owned by The Weather Channel Enterprises Inc., receives floods of visitorsquadruple its normal average page viewswithin two or three days of impending bad weather. Adding to the blizzard, visitors are coming to the site from major portals like Yahoo.com, which license and link to Weather.coms content. With that kind of volatility, IT managers at Weather.com dont need Doppler radar to understand that the site can easily be overwhelmed at any time.
"We have drastic scales here, and the key is we want to provide a very consistent customer experience," said Mark Ryan, chief technology officer of Weather.com, in Atlanta. "Our goal is to have 90 percent of pages download to the end user in 8 seconds or less. To achieve that, we have to scale efficiently."
To reach those kinds of response times and guarantee unforgiving consumers that Weather.com will be online 24 hours a day, Ryan and his team recently revamped the hardware architecture on which the company runs its core Web and application servers. Besides beefing up capacity to withstand well in excess of 30 million daily page views, Ryan and company replaced a few large servers on which it had been running Web applications with distributed clusters of up to 200 midrange servers better able to handle the unpredictable load.
Experts say the unpredictable and growing demands of e-business deployments require companies to evaluate the structure of their Web and application servers to avoid slowdowns, outages and site failures. Meta Group Inc., in Stamford, Conn., recommends that companies build server capacity to handle at least 10 times the predicted normal workloads in the Web environment. The analyst group also says that e-businesses will require application servers to exhibit the same reliability, redundancy and performance demanded of Web servers.
That has certainly become the case at Weather.com. While the sites heaviest load to date was 27 million hits last year during a winter storm, Ryan said his servers can now handle up to 30 million page views a day.
But the site wasnt always able to weather the storm. At the beginning of last year, the company was running the sites back-end architecture on a few very large servers from Sun Microsystems Inc., of Palo Alto, Calif.
Relying on a few large machines created bottlenecksparticularly in data I/Owhen hits unexpectedly rose. Sometimes, download times ballooned to 20 seconds.
In April, Ryan decided to rearchitect his Web and application server platform by moving on to about 200 E4500 midrange servers from Sun running the Solaris operating system. The idea was to spread the load over more machines and take advantage of their increased combined I/O. The new architecture, hosted at two Exodus Communications Inc. Internet data hosting facilities, one in Santa Clara, Calif., and the other in Reston, Va., went live July 1.
"We had the wrong design point," Ryan said. "We discovered that the structure we were using actually caused CPUs to idle."
To make the new architecture fly, Weather.com, with Suns help, had to develop clustering and load balancing software for the Sun servers. Though Sun late last year rolled out its own Sun Cluster software for Solaris, it didnt come in time for the Weather.com project. If there is a failure on one of Ryans servers, the software developed by Sun and Weather.com make sure the Sun cluster will balance the workload across all his servers, enabling users to get the data theyve requested.
Experts say this is a good tactic. Meta Group says e-businesses should employ clustering, load balancing and failover technologies to create a reliable and redundant environment. The analyst group also says the use of technologies to split processing loads across pools of resources is a good idea.
Customizing the weather
Reconfiguring the sun servers wasnt the only change Ryan made to the Weather.com architecture. Last summer, after testing alternatives, Ryan elected to move static content such as images and weather maps off his Sun boxes and on to Intel-based servers running the open-source Linux operating system.
Ryan and his IT team determined that, while some customized Weather. com applications ran well on the transaction-oriented Solaris operating system, serving static content was more efficient on the Intel-based platform. Those pages, representing about 70 percent of the data on the site, are served today by the Apache Web server.
Weather.coms transactional applicationsused to customize content for site visitors, among other thingsremain on the Solaris platform. Those functions represent about 70 percent of the total processing load, as customized user applications require more CPU capacity.
Ryan said he did not consider other operating systems for the transaction-oriented applications because he wanted to draw on the knowledge his staff already had with Solaris and felt comfortable with the operating systems reliability.
Weather.com was able to take advantage of its new architecture late last year when it unveiled a beta site providing new, individualized content to users.
New features, which are expected to increase hit and transaction rates, included severe weather alerts, hourly details on local weather and ski reports. That information is dynamically generated on the fly using information called out of an Oracle Corp. database.
Using IBMs WebSphere application server, also running on Sun Solaris machines, Weather.com is able to request a users ZIP code. Then, using a template, Weather.com dynamically generates meteorological data and other lifestyle informationsuch as for golfers and health plannersfor the users specific location.
In the future, Ryan says Weather. com will build upon its Solaris architecture to bolster its wireless services, as well as to provide more individualized content for users of Palm VII handhelds, cell phones and pagers.
In the meantime, Weather.com is trying to keep on top of escalating hit rate and capacity trends by performing a detailed analysis every couple of weeks. Ryan is keeping a close eye on capacity increases being generated by new wireless features and by partner sites linking to Weather.coms content. As those loads go up, the new architecture allows Ryan to incrementally add servers to his architecture.
And thats a good thing, because predicting Web traffic, like predicting the weather, is not an exact science. With its new server architecture, Weather. com can now prevent a flood of users from taking down the site. And it means that when theres stormy weather, the Weather.com site will continue to shine.