Alot can be learned from those who venture first and take the arrows. Web hosting and Internet data center operations offer the latest case in point. The trailblazers took some fatal hits last year, exemplified by the bankruptcy of Exodus Communications Inc. Now a second wave of settlers in Internet data center territory is trying to avoid the same fate. But this second wave will have access to new and improved weapons -- called blade servers -- to reduce equipment, real estate and operations costs, as well as make the creation of new revenue streams an easier proposition.
With industry analysts like IDC projecting blades will grow to $2.9 billion, or 23 percent, of the server market by 2005, Compaq Computer Corp, Dell Computer Corp., Hewlett-Packard Co., IBM Corp., Sun Microsystems Inc. and newcomers like RLX Technologies Inc. -- widely credited with starting the craze -- and Surgient Networks Inc. are announcing blade products.
"Three years ago, as the switch from relatively small, corporate client/server architectures to large Internet data center architectures began, we saw that it would mean greater volumes of transactions and new security issues, and we saw that Internet data center operators were over-provisioning Internet data center equipment to accommodate that," says Brian Cox, product line manager for HP's business systems and technologies organization. "That led us to question basic server design."
HP announced availability of its first blade server series product in December 2001 and unveiled an alliance with some 30 server software providers. The providers agree to build to the CompactPCI industry standard for a server bus -- the communications fabric shared by the processor, disk drive, memory and input/output (I/O) components in a server. These already are popular among telecom carriers for reliable back-office operations.
Among the first fruits of those partnerships, F5 Networks Inc. announced a server load-balancing product, and SonicWALL Inc. announced a firewall product, both built on HP CompactPCI hardware.
The announcement brought Sun Micro-systems and others out of the woodwork to remind the market of their own blade efforts.
Yet some vendors say the blade may represent the start of a revolution that will include greater computing power, greater I/O, per square foot and per dollar, as well as fully automated loading of servers with the software that actually makes the money.
Shared Costs
|
Older, six-foot rack and pedestal server designs forced operators to buy server capacity ahead of actual customer demand. Subsequent, lower cost 'pizza box' server appliances with much smaller form factors have enabled staged growth, but each new appliance purchase has required service providers to also buy more power and storage and wrestle with a tangle of additional wires.
As a remedy, blade architectures offer a small chassis into which buyers merely can insert an additional, plug-and-play server-on-a-circuit card. They claim a pay-as-you-grow proposition for each Internet data center functional element: web server for serving content, application server for computation and disks for storage. The common chassis enables shared power, wiring, I/O and management components across all the server blades. Alternatively, some manufacturers integrate the I/O into each blade. At his discretion, the Internet data center operator can add components of his choice, including web servers, application servers, storage, I/O or management blades, effectively paying only for the blade and not additional power and other components already in the chassis.
"Flexibility is our main design emphasis, because different blades are best for each type of workload -- application serving vs. stream serving vs. database transactions or security transaction acceleration," Cox explains.
He says he expects HP partners initially will target network edge applications, such as firewalls, virtual private networks, content caches and wireless access protocol gateways, followed by Internet data center infrastructure including content, application and e-mail hosting; streaming; load-balancing; directory services and security management.
Attempting to ride herd on a total package, HP offers data center design consulting, installation and integration, along with bringing its resale channels to bear for software partners. It also packages its HP OpenView management system with tools to enable staging and monitoring of software across multiple Internet data centers.
Buyers will have no easy time sorting through the various vendor claims. Sun, citing its own Netra HA 3.0 hardware and software redundancy, criticizes HP's blade as insufficiently bullet-proof for high-availability service provider operations. HP criticizes RLX, resold by IBM, as inflexible in supporting a variety of workloads. Counter-claims outnumber offers.
Nevertheless, all the proponents agree this fundamental shift in design adds up to lower costs in Internet data center real estate, heating and cooling, wiring and local routing and switching gear, as well as reduced operational costs. HP, for example, claims to require half the floor space of '1U' (two-inch) pizza-box servers with less power consumption, and that a blade can be installed "by the janitor."
Wrong Job
But just how much of a magic bullet do blades represent? According to Nagi Rao, president and CEO of Surgient Networks, they fall short of addressing a tougher issue.
"Servers were never designed for delivery; only for computation," he says. "People have tried to optimize them for I/O and then found bottlenecks in their ability to perform both tasks."
Consequently, Internet data centers have had to use a "Band-Aid," in the form of load balancers, to redirect traffic to the server with the least busy I/O port, he notes, adding that they also have over-provisioned servers, often at a factor of eight to 20 over average loads, to accommodate peak loads for popular web events such as the Victoria's Secret fashion show.
Surgient's eQualibrium architecture and eQ2500 blade product aims to make load balancers obsolete. It specializes in "I/O-intensive" workloads by applying Internet traffic management protocols, including DiffServ and MPLS "and transparently flowing them all the way back to the disk," says Rao.
This enables operators to offer multiple classes of service and give priority treatment to priority users or content, translating to premium-priced services. Other vendors, such as Netscaler Inc. also are applying app-level intelligence to prioritize traffic and fix the server's fundamental I/O burden.
"A data center is a resource to be optimized, very similar to airline seats where the price is based on the customer's judgment of value," Rao says. "The macro impact of this is that we end up helping the cost side through blade densification, but also the revenue side by maximizing the revenue you get out of x square feet by factors of three and higher. We offer real-time yield management to end the era of over-provisioning."
All these approaches could prove complementary, says Rao. "You can take an HP server and tackle computationally intensive apps while we attack I/O-intensive apps."
Internet data center newcomers entering a territory already marked for danger at every turn can only hope that's true.
