Guild Companies

If the only server that you have ever used has run a variant of Microsoft's Windows operating system, then you probably do not know that an operating system can be a lot more reliable than Windows. But many midrange shops do know, and that is why a high-availability solution for Windows, developed by Marathon Technologies and implemented by high-availability specialist Key Information Systems, will probably appeal to the majority of midrange shops that have Windows servers doing at least some of their work.

Marathon, based in Boxborough, Massachusetts, was formed in 1993 by a bunch of fault tolerant computing experts from the former Digital Equipment Corp. who were involved in the development of DEC's VAXft fault tolerant server line. These DEC engineers and marketeers saw, correctly, that Microsoft's Windows NT operating system, which was announced in 1993, was going to need some help if it was going to offer the kind of reliability that midrange Unix operating systems and proprietary operating systems, such as OS/400, offered. The reliability of early releases of Windows was quite bad, and even the later Windows 3.51 and early Windows NT 4.0 releases could only boast about 97 percent availability (that's 263 hours of downtime a year), while midrange machines that competed with Wintel iron could offer 99 percent or higher availability (between 10 and 90 hours a year was typical downtime). Though Microsoft has since done a pretty good job improving the reliability and availability of its Windows NT and Windows 2000 operating systems, the bar has also been raised. Companies that would tolerate 99 percent availability a few years ago want the so-called "five-nines availability" (meaning the machine and its operating system are working 99.999 percent of the time; this equates to five minutes of downtime a year).

This is a lofty goal, and there are a couple of different ways to get such high availability out of Windows, or indeed out of any other type server. While Microsoft and a number of its partners have over the years focused on failover clustering, such as the "Wolfpack" Cluster Services for Windows NT and Windows 2000, Marathon has taken a different approach. Rather than making customers get high availability, which requires them to spread their databases and applications over multiple servers and manage them as distinct units that each do a piece of the processing required in an application, Marathon has created a tightly clustered system comprised of four servers that looks like a single server running a single instance of the Windows operating system, as far as Windows applications are concerned, and all of the high-availability clustering and failover for processors and I/O is completely invisible to those applications. This is true fault tolerance, as opposed to system clustering. The clustering is done at the hardware abstraction layer, and applications do not have to be cluster-aware.

The Marathon solution for making Windows fault tolerant has the unwieldy name of Marathon Assured Availability server. The core program behind this solution is called Endurance. Endurance 6000 runs on Windows 2000, while Endurance 4000 runs on Windows NT. To make a fault tolerant Windows server, you buy two identical Intel-based database/application servers and two identical I/O servers, which are also based in Intel iron. Each of these machines is equipped with a Marathon interface card, and all four cards are cross-coupled to one another and provide the link between the I/O servers and the application/database servers. The I/O servers, which house identical copies of data on disk drives and provide other peripherals, are linked to one another by 10 Mbit, 100 Mbit, or Gigabit Ethernet links. The I/O servers also link to the LAN to hook users into the application/database servers. Each I/O server is attached to each application/database server. The Endurance software runs on all of these machines and presents a single-system image of Windows, with the profile of one application/database server and one I/O server, to applications. Windows is loaded on top of Endurance, and data and applications are loaded on top of that. As the applications run, all processing is performed identically on both sets of machines. If a processor or other component of one of the application/database servers fails, the other server doesn't take over, as in a failover configuration. The second server is actually doing the same work as the primary, in lockstep, instruction by instruction, and it just feeds user screens as if it were the primary, if there is a problem with the primary. Similarly, if an I/O server fails, the other seamlessly picks up the workload. A failed server can be taken offline and repaired, and more importantly, it will resynchronize itself to the working components once it is brought back online.

The basic Marathon offering creates a fault tolerant server with all of its components located in a single data center. But the company has also developed a disaster tolerance configuration called SplitSite, which allows the two halves of the Marathon cluster to be separated by 500 meters, 10 kilometers, or 55 kilometers.

Key Information Systems, an iSeries and Wintel high-availability system integrator based in Woodland Hills, California, has expertise selling and implementing the high-availability software from Vision Solutions for the iSeries. Key was formed in 1999 with the goal of locating and building expertise in various midrange high-availability technologies, and it is one of the primary resellers of Marathon's products in North America. Key is interested in building a business around the Marathon products because more and more midrange shops not only use Windows servers for print, file, and Web serving, but they are using them in mission-critical, distributed applications--often in collaboration with other Unix or OS/400 servers--as well. An application is only as reliable as its most unreliable component, and most midrange shops know that they need to boost the reliability of Windows before they deploy vital applications on it. Even powerful Wintel servers are cheap enough that buying four of them is not that expensive, so a solution like the one Marathon has developed could be a great thing for midrange shops.

Key says it can integrate the Wintel servers and Marathon software and hardware into a fault tolerant Windows cluster for somewhere between $30,000 and $100,000. Right now, the Marathon solution is only available for stand-alone Windows servers, but the engineers at Key are examining the possibility of implementing Endurance on Integrated xSeries Server coprocessors for the iSeries and AS/400 line, as well as for external xSeries machines attached to the iSeries (and using its disk capacity), via the Integrated xSeries Adapters.