of the overall execution -- an amount that we consider to be a statistically insignificant variation. Clearly, any of these systems will provide excellent performance.
Of the remaining systems, the InterServe MP-660 claimed fourth place, taking about 5.5 seconds longer than the ProLiant to complete the 5600-URL test. The Sun Netra and the SGI Origin required nearly twice the time of the other servers, trailing the leader by 26.8 and 30.3 seconds, respectively.
The elapsed time required to complete the Web test is a valid measurement. But equally interesting is how well the server accommodates a rising level of URL requests. All the servers we tested responded similarly to increasing loads. For each 700 additional URL requests, for example, the time to complete increased on average by 3 to 4 seconds.
The AlphaServer again proved to be the performance leader in the database test, besting its nearest competitors by 25 perce
nt in both single and queued query conditions. Second place was split among the NT/Intel-based servers: the LX Pro, the ProLiant, and the InterServe MP-660. Not unexpectedly, all three turned in nearly identical times in the database test under all loads. The Netra and Origin 200 posted the slowest times when asked to perform four or more database queries. To process 12 queries, the Origin 200 required over 367 seconds -- 86 percent longer than the AlphaServer and 40 percent longer than the Wintel systems.
The database test measures both computation speed and, indirectly, the overhead associated with each system's SMP implementation. With four CPUs available, we expected that the time to process from one to four requests would be fairly close, with all tasks able to run simultaneously. Generally, this was true. However, the Origin 200 was the exception, with times increasing from a speedy 69 seconds for a single task -- second only to the AlphaServer's 51 seconds -- to the slowest response of 123 second
s for four tasks.
As the number of tasks was increased to eight and 12, the time required to complete approximately doubled and tripled for each system, respectively. Under heavy loads, both the MP-660 and the Netra i 4000 dropped behind the rest of the pack. A time-out in the Netscape LiveWire program expired before the Sun system could complete processing 12 queries, so we were unable to record a finish time.
NT Comes of Age
Windows NT has suffered from a second-class image when compared to Unix servers. And traditionally, differences in OS and server software have made it difficult to compare the two. But using Netscape Enterprise Server as a common test platform allowed us to minimize the distinction and concentrate on the effect of the processor architecture and the OS on overall system performance.
When we pitted these machines head to head, we were somewhat pleasantly surprised to find that the NT/Pentium Pro combination was potent competition for the Unix systems.
After years of playing the bridesmaid to RISC/Unix, Wintel boxes are showing that they can hold their own in the intranet server field.
The three Wintel servers we tested in this Lab Report are all powerful, sporting quad processors and wide network bandwidth. Each is quite capable of running your intranet Web service with plenty of muscle left over for mail and other background tasks. And the combination of processor power, plenty of RAM, and huge disk arrays equips them to speed through large database tasks with capacity in reserve.
Other Considerations
With performance between the systems comparable, price will take on a more significant role in any purchase decision. But there are other factors, both subjective and objective, that you should take into account when selecting one server over another.
Before adding a Unix server to a predominantly Windows NT network, for example, you'll have to weigh how much expertise your company has with that particular brand of Unix,
RISC architecture, and system vendor. Acquiring operational expertise can be an expensive proposition, especially when you consider the investment you'll have to make in setting up, tuning, and maintaining the system. You may also have to spring for a new suite of development tools and applications.
On the other hand, even dedicated Unix shops should consider the less-expensive Wintel servers as alternatives to new Unix servers. Chances are good that some form of Windows already has a significant presence on your network in at least a client role. In that case, making the transition to NT Server isn't such a big leap.
Reliability is also something that's difficult to benchmark. But how valuable is a really fast server that doesn't stay up when you need it? Characteristics that contribute to reliability include the ability to swap failed components, including power supplies and hard disks, without bringing the system down. And even if you can't swap things like CPU, RAM, and network cards, being a
ble to go inside the box and easily replace a bad component is important, because you know you'll have to do it someday.
Best Overall: Intranet Servers
