THE H-REPORT

Web Server Technology Becomes A Team Sport

by Christine Hudgins-Bonafield

There's no such thing as a Web server. At an increasing number of sites, the need for something bigger/faster/better has fashioned the single server into a server farm. And this new corporate foyer has become so critical that its architects are encasing fault-tolerant database architectures within fault-tolerant Web server designs.

One reason, pointed out by Forrester Research, is that half of all sites plan to provide transactional servers soon, while 22 percent already provide them. If a buyer can't get to one site to buy a product, he or she will simply chalk up a mental black mark and go on to the next.

That means high-performance Web servers are becoming mission critical and highly competitive, with some users comparing their network equipment and operations to those of banking networks. The days of Web technology giveaways are shifting to serious business at serious prices. Users want raw speed; hardware reliability lasting months, not weeks; OS-tuned Web software that avoids reboots caused by long-term memory allocation and release problems that slow systems; the ability to unplug a machine in a cluster without downtime; the ability to delegate all but mission-critical Web server functions to offline machines; automated mirroring and caching of server content to other sites; content designed for perform ance; and the kind of self-diagnosis found today on mainframes.

Here's our best speculation on some of the things users will get by year's end:
· Microsoft will augment its overarching WolfPack effort to extend high availability to NT (this will be accomplis hed using volume management technology, already OEMed by Sun, Digital Equipment Corp. and many others, through agreements with software supplier Veritas);
· IBM will announce new object-oriented technology to speed access to databases interconnected to Web sites through Web-based caching techniques;
· Several companies, including IBM, will announce technologies to let Web servers transparently add and drop machines from multiple vendors in a Web cluster;
· Several of the most popular sites, like Digital's AltaVista search service, will be mirrored in multiple locations to improve performance.

Problems Abound While it's often difficult to sort out whether performance problems are caused by the Internet itself, the pipe to the server or the server architecture, Forrester's John Robb suggests that server architectures are largely to blame.

And while Web servers can look a lot like any other server, they do present their own performance and availability issues. Michael Peterson, president of Strategic Research Corp., Santa Barbara, Calif., says that the average Unix server is down about two hours per week for scheduled or unscheduled maintenance. While this may be tolerable on a corporate intranet providing a company phone directory, it can raise serious problems for a multiserver Internet site demanding high availability.

Also, browsers cache IP addresses, so users may be directed to a server that is down. If the IP address that is cached moves round-robin on a server farm to handle load balancing, the sequence can result in the toasting of a second operating server. Much of the problem also lies in the dynamic nature of Web traffic. While those building an intranet may be able to calc ulate an expected maximum number of users simultaneously accessing a server, the same is not true of the Internet. Concurrent access requirements are unpredictable and often grow at incredible rates, leading many servers to be highly tuned for I/O. The need to mai ntain static information--like passwords and database information--behind a server that may go down introduces yet another performance and reliability challenge.

Scott Penberthy, development manager for the core server software being used by IBM for its Olympics server, says some servers that offer a lot of interactive applications tend to take too much time per hit, while turning away other connections. However, if the page is being delivered right out of memory, the OS can choke because it's handling too many connections (more than 1,000 or 2,000) on a given node.

A Reliable Winged Pig All of this means that down on the server farm there's a requirement for scalability, speed and reliability. And the solutions have primarily been to build your own, but Web requirements have now become a major driver for long-awaited client/server technology. So commercial products are starting to emerge.

D. Navin-Chandra, chief scientist at I ndustry.Net, says his company has been forced to rely on server technology customized in-house to achieve high performance and persistence. But he believes Industry.Net shouldn't have to be in the business of developing or modifying Web server software. What he'd like to see is an off-the-shelf product that takes "a bunch of disk arrays and a cluster of, say, a dozen Web server machines and has them all working in tandem as one Web server--so that any hardware item can be randomly taken offline without a service outage." Today, he says, vendors provide "some proxy methods, round-robin and some load sharing, but no reliable plug-and-play way of handling the loads in a smart way."

Of course, clustering technologies are nothing new. Digita l and IBM have been at it for years, and just about every other major server provider has put clustering on the fast track. Compaq and Hewlett-Packard also expect to up the performance ante with quad Pentium Pro 200-MHz servers. The problem for c lustering, especially fault-tolerant clustering, is that it still tends to be a pricey single-vendor solution. And depending on how you define fault-tolerant, it may not be available at all.

Veritas Extends High Availability Many users would like to tap existing and even multivendor equipment for server farms. Peterson believes that one of the main avenues for achieving this goal will be through a traditional OEM company that is now shrink-wrapping its technologies: Veritas.

Veritas has sold its file and volume management technologies to almost all of the primary Unix providers as well as fault tolerance experts like Tandem. It is also extending its technology beyond Unix to Microsoft's NT and is one of the few Microsoft partners with access to NT source code. Company officials say they cannot comment on whether Microsoft will incorporate Veritas' technology into a fault-tolerant version of NT, but we think a match is in the making.

Peterson says Veritas' importance l ies with its ability to manage storage systems. Up to 60 percent of downtime is caused by storage-related failures, including lack of disk space, says Peterson. "If you look at the top two storage management processes that contribute the most to network availability, at the top is the hashing and replication of file systems and the second is server marrying. Veritas offers both. It has the top two products in this [rather small] field."

Roger Klorese, manager of marketing programs for Veritas, says the company expected to be beta testing its First Watch product (which operates in a pair of servers to provide either hot standby fail-over or to reconfigure the pair to exclude the downed server) on NT by third quarter. It planned to roll out its Volume Manager (technology to build groups of disk storage that can be physically moved between machines or failed over) on NT by early 1997.

Klorese says NT has been able to protect data through mirroring but the reboot required in the p rocess inhibits availability. He also says that NT needs to become scalable beyond four to eight CPUs--and as it does so, managing the disks hanging off those boxes will become more important. Today, Veritas can manage storage on 30-processor Sequent or Sun boxes as well as Cray machines with 64 processors.

Veritas is also developing clustered file systems, for which it has a prototype, but the actual product isn't expected to be available until early next year. Automated management (such as policy-based methods for growing a file system) is expected in about the same time frame, although Veritas can already perform management tasks like reconciling snapshots of data before and after a backup to rapidly present a consistent image.

Veritas also planned to be shipping a Web Edition, based on FirstWatch, Volume Manager and the Veritas File System for Sun's Solaris running Netscape, NCSA or Apache servers by this month. That edition includes agents on each machine to monitor the status of an other and bring itself up as the Web server with a spoofed IP address, if a failure occurs.

The Good, the Bad, the Ugly Are there servers or server components to be avoided at industrial-strength Web sites? Peterson says 25 percent of Web sites are run on workstations that users have turned into servers--and that's to be avoided, primarily because of insufficient bus speeds. He also believes NT is more appropriate for intranets than Internet sites. There's a reason why approximately 60 percent of Internet sites are run on Unix, says Peterson.

However, Navin says he invested in a $40,000 IRIX-based Silicon graphics machine at Carnegie Mellon University (where he also holds a teaching position) only to find it great for content creation but lacking in long-term reliability as a Web server. He also thinks NT isn't yet up to the task, although he's keen on the price-performance of Digital's Alpha machines as NT servers. And while Apple machines certainly have a presence on the Inte rnet, they're often discounted for the largest, clustered server sites. IBM, of course, sees some real advantages in tapping the mainframe for the Internet.

As for the actual server software, Forrester estimates that 80 percent of the professionally managed sites use Netscape servers, 37 percent use shareware and 5 percent use Microsoft servers (some sites use more than one). Navin likes the fast CGI capabilities in OpenMarket's servers and the access to code in Apache shareware, but adds that server functionality is moving so quickly he'd have to flip a coin to try to make a decision today.

Christine Hudgins-Bonafield can be reached at cbonafield@nwc.com.

Updated July 8, 1996