Dynamic Resource Sharing Transforms Data Center Design

While traditional telephony carriers may engineer their switches and trunking capacity with something short of Mother's Day peaks in mind, managed hosting providers find they must engineer for a greater number and variety of such peak events, whether widely publicized online lingerie showcases, corporate events or unforeseen national emergencies.

So, managed commerce providers like Digital River Inc. and a host of Internet data center hardware and software vendors are attempting to draft a new map for engineering facilities.

Rather than adding server, database and other capacity on a customer-by-customer basis, the new map is framed by the goal of allocating pools of shared computing power, bandwidth and other resources in real time and near real time in response to customers' changing needs.

"We've built a system that allows dynamic reconfiguration of resources," says Marty Boss, chief architect and vice president of information systems for Digital River, which provides managed hosting of compute-intensive commerce services for online stores run by customers including H&R Block Inc., Major League Baseball's MLB Advanced Media, L.P. and security software seller Symantec Corp.

Although Digital River can plan for a three-times-normal load on H&R Block's databases, application servers and web servers on April 15, or on MLB's store at World Series time, sudden rushes on Symantec's online store for the latest Internet virus fix are less predictable.

Foreseeing this scenario, the company began three years ago to design its infrastructure and resource provisioning systems around dynamic reallocation. That work led to a design that manages groups of database, application and web servers in "pods" that may be shared by multiple online stores. If, on Pod 1, traffic in and out of one of those stores passes a threshold trigger, the system automatically shifts the content and transaction processing of that store to a burst capacity pod, leaving the others on Pod 1 to operate at normal loads.

Now, Boss says, as Digital River grows, rather than sizing its Sun Microsystems Inc. servers and Oracle Corp. databases and applications on a 'one-off' basis for each customer alone, that sizing is done with aggregate peak loads in mind.

The provisioning system, built in-house, was integrated with Sun's Solaris operating system, which allows dynamic partitioning of processors and assets across multiple servers. "With Digital River, most of the applications may not normally scale beyond two central processors," says Chris Kruell, group marketing manager of enterprise system products for Sun. "For companies with larger scale, compute-intensive applications, consolidation of servers becomes critical, in terms of parts, management systems and expertise."

Boss suggests Digital River's operations generally fit that description. "All we do is commerce and transaction management, serving 30 to 50 unique pages per second, each generated dynamically by our backend," he says. This adds up to a pressing need for point-and-click management of capacity -- including processors, memory, databases, storage and input/output -- across its infrastructure as a single, consolidated, fungible entity.

Off-the-shelf products targeted to this goal are emerging from vendors including Ensim Corp., Inkra Networks, Sphera Corp. and SWsoft.

SWsoft's HSPcomplete partitioning solution, for example, allows hosters to offer about 800 virtual private servers per physical server, with each Virtual Environment acting like a single instance of a Linux operating system. The company says this is ideal for a hoster's growth down-market into medium-sized companies with limited budgets.

In sizing capacity for a new customer, a provider must go beyond forecasting the number of hits, transactions and downloads per day to further assess and reassess which functions are growing over time, Boss says.

"One client may have lots of hits but few downloads, so I may only need more bandwidth and a front-end server, while another may have a lot of transactions, requiring more backend database only," he says. "Some sites may have more browsing than closed sales, and typically, the higher the close ratio, the more resources I'll need for credit checks, messages to third-party fulfillment partners and other transactions."

Layers Of Sharing

Vendors of all stripes are attempting to provide out-of-the-box shared resource allocation.

To help Internet data centers achieve on-demand scale, Hewlett-Packard Co., IBM Corp., Sun and other vendors are racing toward server blade architectures, with plug-and-play circuit cards designed for grow-as-you-go capacity increases (see February 2002, xchange page 30). But that revolution, represents only one layer of innovation around just-in-time architectures.

According to Dave Roberts, co-founder and vice president of marketing for Inkra, provider of service activation systems designed to coordinate multiple Internet data center systems. "Blade servers, shared pools of virtualized storage and of network capacity are all going through the same shifts for the same reasons -- virtualization of physical capacity."

The main goal, through automation of sharing, is to reduce "humans per square foot, making services more productized, repeatable and automated," Roberts says. In that scenario, the provider can offer "premium overflow service, allowing a customer using two servers to overflow to server number three, and the premium helps pay for that third server."

Inkra hopes to supplant the currently common "brute force" practices of racking and unracking, configuring and reconfiguring -- anything requiring a human with a screwdriver or software code assembler -- because those processes constitute "a direct hit on profitability and make up-selling and cross-selling additional services more difficult," Roberts says. The company expects to release its first product in the first half of this year.

With dynamic partitioning and sharing technologies emerging at multiple layers, from app-intelligent packet switches and load balancers to blade servers, partitioned operating software and dynamic resource management systems, Internet data center operators may be approaching what Roberts says is the "final solution -- a lights-out, data-center-on-demand solution, with fewer operations staff, cheaper real estate and an architecture built around overflow premium capacity."

Share this article: Email, Slashdot, Digg, Del.icio.us, Yahoo!MyWeb, Windows Live Favorites, Furl
Add this article feed to: RSS, My Yahoo, Newsgator, Bloglines