Sold on the new data center concept

Burlington Coat Factory is investing its IT future on a grid-based, virtualized architecture.

Burlington Coat Factory CIO Mike Prince has ridden many a technology wave in his 21 years with the company. But none has inspired the awe he feels about the new data center. Automated provisioning, grid computing, high-speed system interconnection, open source systems, virtualization: “This approach is so rich and feature-full, it’s overwhelming to think of how we’re going to exploit it all,” he says.

But exploit it, Burlington Coat Factory will.

Never one to shy away from a challenge, Prince is piecing together the latest data center technologies to create a sophisticated on-demand architecture with extreme economy – processing power that previously would have cost $1 million will cost only $100,000, he says. To do so, Prince is picking technologies from industry stalwarts and start-ups alike.

“We didn’t have to read the tea leaves to know the best technology out there for a new architecture,” Prince says. Call it grid, as does Oracle, or on-demand, like IBM, he adds, “I absolutely believe in the basic concept of binding together slower systems by hardware and networking so they can be used in parallel to provide computing resources and in so doing creating a highly scalable and reliable environment.”

Burlington Coat Factory is not just changing out the operating system, database and systems hardware used in the data center, it’s undertaking this tri-level migration simultaneously. Under the new data center architecture, out goes the Dynix/ptx Unix variant and Oracle 8i database on IBM Sequent servers; in comes SuSE Linux (Novell) and Oracle 9i and 10g on Intel-based IBM xSeries machines. Topping off this infrastructure mix are new data center products such as Cloverleaf Communications’ disk virtualization technology, PolyServe’s file management systems for clusters, Topspin Communications’ InfiniBand-based server switches with server virtualization software, and Vieo’s application infrastructure management appliance.

These new data center plans mushroomed once Burlington Coat Factory learned in 2001 that IBM was closing down the Dynix/ptx line, and that Oracle was backing off support. First, the company quickly settled on SuSE Linux as its new operating system. But because Burlington Coat Factory could neither run Linux nor the latest Oracle databases on its old hardware, nor put its old database on the new platform, hardware and database change-out became equally imperative. The only sensible choices were IBM’s Intel-based xSeries servers and Oracle databases running over Linux, Prince says.

Even with the other changes, the operating system decision was easy, given the company’s familiarity with, and advocacy of, Linux. Burlington Coat Factory has used Linux for its point-of-sale and backroom retail operations since 1999. “We know it. We like it. We believe in it,” Prince says of the open source code already deployed on about 7,000 POS systems and small computers.

Burlington Coat Factory began casually investigating Linux in the data center in 2002 and started a serious migration attempt in early 2003. But that effort lost full steam by August 2003, when developers realized “too many pieces of the stack weren’t ready,” Prince says. The IT staff said a Linux-based data center couldn’t quite handle the company’s high-season needs. At the time, SuSE Linux couldn’t capably handle the company’s disk access requirements. The operating system could perform hundreds of logical unit number (LUN) lookups, but Burlington Coat Factory needed to do thousands. Plus, the Cloverleaf disk virtualization technology wasn’t ready to move out of beta-test mode, adds Prince, noting that the retailer made it through last year’s holiday shopping madness with the help of loaned Sequent gear from IBM.

Prince foresees no such problems this year. The technology has matured to the point that by Labor Day, the $2.7 billion Burlington, N.J., retailer expects to have its eight most important databases – those used for merchandising information – running in an Oracle 9i- or 10g-based grid, he says. “The hardware is in place, the storage provisioned, and everything looks extremely good in testing and benchmarking,” he says.

He notes, too, that because Burlington Coat Factory is changing out its operating system, hardware and databases all at once, it has taken an ultra-rigorous approach to integration and testing. “We’ve been pushing our project management and quality assurance to their limits,” Prince says.

For example, developers have been “waling” on a full-blown replica of the men’s outerwear database – one of the eight merchandising databases – “and it’s looking very good in terms of performance,” Prince says.

The company can tap into available computing resources as needed from any number of the distributed systems tied together in the Oracle 10g grid. If Oracle10g continues proving successful in test mode, then Burlington Coat Factory likely will skip over Oracle 9i for this men’s outerwear and other critical merchandising databases and move right into the 10g-based grid.

Still, Burlington Coat Factory continues putting Oracle 9i through its paces, both in test and production scenarios. For example, a stored-value-card application, which keeps track of the value owed to customers for gift cards and cards issued for store credits, is powered by an Oracle 9i Real Application Cluster (RAC) in the new data center. Oracle’s 9i RAC lets users run databases across multiple servers, providing load balancing, fail-over support and scalability.

No matter whether 9i- or 10g-based, Burlington Coat Factory’s new database and application server clusters will use InfiniBand for high-speed interconnection. With InfiniBand, the company will net even greater performance improvements than it can achieve by moving off the old large-scale Unix systems alone. “We hope to double our Oracle application performance,” says John Decatur, a systems specialist with Burlington Coat Factory.

Using Topspin’s InfiniBand-based server switches in test mode, the company has interconnected the clustered xSeries database and application servers, the enterprise LAN servers and the Fibre Channel storage-area network over a 10G-bit/sec fabric, Decatur says. “With Topspin, we can bring the network and the storage right to the switch, which makes for a much simpler model” than the conventional network architecture that requires a separate interface for each system interconnected, he says.

With Topspin’s new VFrame server virtualization software, Burlington Coat Factory can program the server switches with policies it needs to create virtual servers out of shared computing, storage and network resources. The company has tested the InfiniBand capabilities of the Topspin products for about six months, but only recently began trying out the VFrame capabilities, Decatur says.

However, the VFrame virtualization software brings out one of the biggest challenges developers have found in architecting for the new data center – that being, trying to figure out what vendor to choose for which function, Prince says. For example, Burlington Coat Factory has been testing – and liking – Cloverleaf’s disk virtualization technology for provisioning storage resources (at 60T bytes today and climbing rapidly, Decatur says). But VFrame, although primarily for server virtualization, provides some functionality for virtualizing storage, too.

“Topspin is providing InfiniBand resources for core networking and storage. But it also provides the capability to provision servers, pulling together not just the binaries for the operating system and starting them up but also providing the correct network identity and visibility into storage-area networks,” he says. “And the LUNs that all these servers see? They are virtual LUNs being created by Cloverleaf.”

And the overlap doesn’t stop there. Burlington Coat Factory can use Oracle for some storage provisioning and also plans to test Vieo’s application infrastructure management appliance for its ability to command provisioning should service levels fall below expectations. Then in the storage layer, there’s the PolyServe software the company plans to use for provisioning a clustered file system on the clustered servers.

“The most difficult thing, having grappled with this for a year, is that we have several good choices on how to play this . . . We see a tremendous amount of synergy and the potential to automate the data center like never before. . . . But we’re looking at five or six vendors in the stack – that’s a lot,” Prince says, especially because these aren’t just vendors providing nice-to-have features but major functions of the new data center architecture. The good news is that no matter the choice of vendor – or combination of vendors – deploying server and storage virtualization will increase computing power and performance, he adds.

More efficient use of computing resources also will let Burlington Coat Factory undertake resource-intensive analysis projects it previously shied away from because processing requirements were too high, Prince says. With reliability of the new architecture expected at near 100%, the company should be able to let applications run at the retail outlets pull information out of centralized databases. Previously, Burlington Coat Factory did not support such interactivity between the stores and the data center. Prince explains: “We didn’t want an operation at a cash register interrupted by a computer disruption at the data center.”

But Burlington Coat Factory certainly is proving that change is good. As Prince says, “We have painstakingly proven that [the new data center] approach works. This is not vaporous.”