InformationWeek.com

Welcome Guest. | Log In| Register | Membership Benefits

November 16, 1998

Server Evolution

NT and Unix vendors plan to pump up power, add features

By Mary Hayes and Martin J. Garvey

Illustration by Michael Aveto/SIS

erver hardware technology has evolved at an awesome rate in the past few years, but even that breakneck pace hasn't kept up with IT managers' demands for powerful, reliable, manageable, and affordable systems. Fortunately, the future has never looked brighter. In the next several years, the evolution of server hardware for the most commonly used operating systems is about to shift into overdrive.

On the Intel platform, more powerful chips and improved system de- signs will firmly plant these "commodity" servers in the midrange market. And RISC-based systems, which proliferate in the midrange and high-end spaces, are approaching mainframe-level status in terms of power, reliability, and avail- ability features. IT managers are likely to be able to get more function and capacity for less money in the coming years.

Such advancements will occur because the server market is extremely lucrative for hardware vendors. The $32.2 billion market, according to Dataquest, is expected to grow at an annual rate of 15% over the next five years. Customers, of course, have plenty to tell these vendors about the technology improvements they need to make to earn, or keep, their business.

At United Airlines Information Services, RISC-based Unix servers provide a platform that's more flexible and scalable than mainframe systems for data warehouse and online analytical processing applications. But Ajay Singh, manager of research and development, needs more. Singh wants to integrate three critical applications that reside on separate, networked boxes--aircraft scheduling, revenue management, and ticket pricing--on one Unix system. The Elk Grove, Ill., company wants one server that offers the power and capacity to handle these processes, and the automated management features to support them, for a couple of reasons: It costs less to run and manage one server, and performance improves because no input/output processes need to occur among multiple systems.

"As opposed to solving things piecemeal, we'd like to solve them together," Singh says. "In essence, we want an integrated platform for the entire planning process." But Singh says that compared to mainframes, United's massively parallel IBM RS/6000 SP Unix systems fall a little short in areas beyond flexibility and scalability. Today, none would be capable, on its own, of processing and summarizing the increasing volumes of data that would result by consolidating the applications, while retaining acceptable performance for users.

Meanwhile, there's much to be desired in the reliability of Intel-based systems, users say. Rumpke Consolidated Cos., a Cincinnati waste hauler with 33 sites nationwide, uses Intel servers running Windows NT for payroll and accounts payable applications, while retaining its core financial apps on two IBM AS/400 midrange servers. The company says it chose the NT platform for its low cost, compatibility with client PCs, and management software that supplier IBM developed for its Netfinity NT line.

Those advantages have led Rumpke to consider developing routing applications for the NT/Intel systems, but the company still has concerns about trusting critical applications to the platform. "The systems are getting close to providing the type of performance that's necessary, but what's most important is reliability," says IS manager Mike Wilson. A failed server could back up the company's trucks--and every minute a truck isn't hauling waste, Rumpke isn't earning revenue.

Vendors are developing technologies to address the concerns of customers such as Singh and Wilson. Some of the efforts involve improvements to proprietary Unix operating systems and RISC-based hardware, and some are more dependent on Microsoft, which is expected to deliver a more reliable and scalable operating system with Windows 2000, formerly NT 5.0, in late 1999. But even so, vendors building Intel-based systems plan hardware advancements to improve the performance of NT today, and deliver a one-two punch once Windows 2000 is available.

Premillennium Progress
Advancements coming in the next six months include new chip and memory designs that some vendors say will offer improvements over symmetric multiprocessor systems for midrange to high-end servers, and greater scalability and clustering options for Intel servers.

By year's end, for instance, Hewlett-Packard says it will complete the integration of Non-Uniform Memory Access technology into its HP-UX operating system, and make that technology available in its high-end V-Class commercial system.

HP and other NUMA Unix hardware vendors, such as Sequent Computer Systems, say NUMA will improve the scalability of any system, using a cross-bar interconnect design to free memory from sharing a system bus with processors, thus overcoming latency as more processors are added. Overcoming latency is an attractive capability, says Singh. "For some queries [with United's OLAP applications], that means a response in five minutes instead of an hour," he says.

Boosts in chip power will continue to propel the performance of RISC-based servers, allowing them to retain their leadership as the open platform for database and data warehousing applications by reaching a record number of transactions per second. Next year, Sun Microsystems debuts its next-generation Ultra Sparc 3 chip, with clock speeds increasing from 400 MHz to 650 MHz, while IBM's PowerPC chip will go from 262 MHz to 450 MHz. In early 1999, HP will introduce a new PA-RISC chip that will have internal clock speeds up to 440 MHz. HP says the yet-unnamed chip, along with other system design improvements, could boost performance by as much as 70% over current V-Class systems.

Vendors will also be packing more RISC chips into systems. IBM will double the number of chips per system to 24, and improve its cross-bar interconnect so that data travels at 22 Gbps, up from 11 Gbps. While Sun already has a 64-CPU server, vendors, including HP, will launch their own 64-CPU machines in 1999.

A few Intel vendors, such as Sequent and Data General, can already fit dozens of processors into a system. But those are highly specialized units that can cost hundreds of thousands of dollars. In the first half of 1999, companies that deal in the less-expensive Intel commodity server market will begin to include eight-processor servers as a standard option among their offerings.

That would suit Rumpke's Wilson just fine, because he also cites increased scalability as a requirement before moving AS/400 applications to NT servers. It will also help that Intel will begin to churn out upgrades to chips designed specifically for servers, beginning with the 450-MHz Xeon in early 1999.

But even more interesting is the work that vendors are doing with system partitioning and server consolidation. "With every Unix vendor, the discussion centers around partitions," says Jonathan Eunice, an analyst with Illuminata Inc. Partitions let customers separate application resources on each system, so that they can adjust resources to accommodate added or expanded applications. In fact, without partitioning there's little call for servers with 64 or more processors, because few single applications require the use of that many CPUs.

Partitioning, which originated as a mainframe technology, will let IT managers integrate complementary applications on a single server--rather than pay the hardware costs, use up the space, and deal with the integration issues of crowding a number of servers into the data center. The idea appeals to many, including Dayton Hudson Corp., a retailer in Minneapolis. "We've got to get out of the mode where every application requires its own server," says Mike Thyken, director of technical services, who manages 200 servers. "We need to get to a tenth of the number of servers we have if we want to be cost-competitive."

Vendors such as Compaq, HP, IBM, and Sun are working on partitioning technologies for Unix systems this year, beginning with hardware partitioning and evolving into full workload management via software within the next few years. Workload management will automate the balance of resources among partitions, so that systems won't have to be shut down for resource allocation. Sequent, in fact, has promised this for early 1999.

Singh of the United Airlines unit says vendors are on track: "We want dynamic work- load management with lots of horsepower and flexibility to use [servers] for multiple processes."

Clustering is a big frontier for both the RISC and Intel vendors, improving system availability and, in some cases, performance. Sun will offer enhancements to its Solaris Full Moon clustering technology for its Sparc-based systems next year, including system failover of up to four nodes. And although NT/Intel vendors won't offer operating system-based clustering to support all types of applications until clustering for Windows 2000 is available, suppliers are jumping headfirst into parallel database clustering.

Software vendors such as Oracle and Microsoft now offer "parallel" versions of their databases that run on two or more servers, to provide customers with more reliable and scalable databases on NT. Compaq plans clusters of Oracle Parallel Server of up to six nodes beginning later this year. IBM, meanwhile, plans to certify eight- to 16-node clusters for OPS or DB2 on its NT/Intel systems within a year.

2000 And Beyond
Sometime in 2000, Intel commodity hardware vendors predict that their systems will begin to meet the midrange server market head-on. New developments expected to start to converge at the turn of the century include the availability of 64-bit processors from Intel and Windows 2000 from Microsoft, and the implementation of high-end system scalability, availability, and reliability features onto the platform.

For instance, some vendors plan to implement the cross-bar interconnect design in their Intel systems, as well as in their RISC machines. HP says SMP designs (including its own systems) eventually won't be able to support technologies such as Fibre Channel storage and Gigabyte Ethernet switch- ing, and it will likely adopt a cross-bar interconnect design throughout its product line.

Also in 2001, a new peripheral interconnect design will replace Intel's PCI bus, which is limited by how fast it can transmit data between servers and peripherals. Intel, Compaq, and IBM are working separately on "switched fabric" channels for Intel and RISC systems.

As Intel's 64-bit CPUs appear in 2000 and 2001, some companies plan to deliver application partitioning on eight-processor NT/Intel servers, similar to that planned for Unix/RISC systems. Intel server vendors say their customers are also asking for solutions that will allow them to consolidate applications, improving resource manageability by limiting, as much as possible, hardware to one server, one set of hard disks, and one set of network interface cards.

Another concept that will pick up speed in 2000 is the idea of a modular server. Vendors such as Dell Computer and HP are investing heavily in the concept, in which hardware components exist as separate modules: There will be a separate I/O card hotel and CPU module. Each module could be trouble- shot and replaced independent of other components, which would save companies money in hardware costs and give them more flexibility in scaling systems.

But one of the most important goals of NT/Intel vendors is to guarantee that systems won't go down for more than a few minutes per year, if at all. The goal of IBM's On Forever project, for instance, is to guarantee that a system will never crash, and that all adjustments can be made while a system is running. Most vendors already offer the ability to add peripheral components to a running system, such as storage controllers and PCI cards. But IBM expects to be able to let customers "hot-plug" processors and memory to its NT/Intel systems, too.

Meanwhile, the Unix/RISC camp will move forward in 2000. At that time, Illuminata's Eunice says that most Unix-based servers will be able to offer about 80% of the performance, availability, and reliability of mainframe systems, vs. 40% to 50% through 1999. Vendors say partitioning will be more common in 2000, and companies are looking at management software to monitor the efficiency of partitions.

Ultimately, these improvements should deliver what customers want. Whether it's an Intel- or RISC-based system, says Singh of the United Airlines unit, "We need it to be scalable and we'll need more performance for larger problems."

Illustration by Michael Aveto/SIS