Kernel Slander

Will Profusion hurt or help the NT kernel's reputation for poor scalability?

The weather in Colorado can change three or four times a day, and rainstorms sweep in quickly and leave just as fast. Sometimes after a storm, a double rainbow appears. These displays usually remind people about the legend that tells of a leprechaun sitting on a pot of gold at the end of the rainbow. Windows NT scalability is a lot like that mythical pot of gold—just when you think you're about to locate the treasure, the rainbow disappears.

Reports of poor or nonexistent NT scalability have plagued Microsoft since the advent of the OS. We've conducted testing in the Windows NT Magazine Lab with NT Server 4.0 on 8-way servers. (To read about this testing, see John Enck, "8-Way Scalability," September 1998.) Single- to dual-processor and dual- to quad-processor increases are possible, but making advances in 6-way and 8-way processor performance with NT has been roughly analogous to chasing the pot of gold at the rainbow's end. Enterprises need the speed that 8-way systems can achieve, but until recently, NT hasn't scaled adequately on 8-way systems. With Windows 2000's (Win2K's) release on the horizon, Microsoft is claiming that the OS's scalability will increase. This claim will either raise hopes or draw skepticism from users, depending on their experience with previous promises of scalability.

Although Microsoft has taken a lot of heat for NT's inability to scale, another culprit has stayed below the media's radar. Intel produces 90 percent of the processors on the market, and Intel processors run NT on most 4-way, 6-way, and 8-way servers. Although many might argue that NT's kernel is a scalability bottleneck, accusing the NT kernel without examining other likely suspects is unfair. The system bus is one of those suspects, yet the system bus is a primary reason for a lot of the excitement that surrounds Intel's release of the Profusion chipset. (For information about Profusion, see Tao Zhou, "Profusion Architecture," page 60.)

Although we don't want to let Microsoft completely off the hook, Intel should be under close scrutiny as well. For example, the Alpha processor continues to bury the Intel chip in scalability and performance tests, and the Alpha has set many performance world records over the past year. Unfortunately, vendors build few applications for the Alpha processor that appeal to widespread use, and Alpha chips are more expensive than Intel chips. Unless you have a narrow performance goal and cost isn't a problem, you're not likely to use an Alpha processor.

The Profusion chipset might let Intel and the system bus off the hook in the scalability debate. Then, if the system bus is proven innocent, Microsoft will carry the burden of proof that Win2K and NT truly scale. We expect Win2K and NT to get a performance boost from Profusion and the 8-way servers that use the Profusion chipset. Microsoft claims that Win2K Advanced Server (Win2K AS) scales to 8 processors and Win2K Datacenter Server (Datacenter) scales to 16 processors.

Stay tuned for the Lab's upcoming coverage of 8-way Profusion servers that run on Win2K and NT. Will Profusion bring an end to kernel slander or only fuel the fire? We're as anxious to find out as you are.