Windows NT vs. Unix: The Chase Is On!
For scalability testing, we generated databases of increasing size on the Solaris platform (for consistency, since Solaris didn't support discovery of IPX STATIONS on the network and the NT version did) and loaded them on both servers. To normalize GUI performance, we set both displays to approximately 1,100x900 pixels at an 8-bit depth.
Our first impression of the NT version of SPECTRUM was that the interface appeared very sluggish--a common, though misleading, finding in qualitative OS performance tests. Without probing further, it would be easy to accuse the NT version of being slow compared with the Solaris code. Because of the way the app was ported (the NT version of SpectroGRAPH runs through a local X Windows server, not natively through the NT GDI), the Sol
aris version's interface was only slightly faster at drawing windows than the NT version. However, all processor-intensive database and polling operations involving the SpectroSERVER, such as opening an object view and querying performance information, were similar across both platforms.
Once w
e started drilling down further into the application, we were surprised to find performance--particularly at the server level--to be very similar. We started small, generating a map of routers throughout the Syracuse campus. As expected in this control test, we generated and loaded a relatively small database of just over 1,000 models and about 500 polled objects onto both platforms. As expected, we could not find any appreciable load on either server and interactive performance--such as querying a node's performance or configuration--was comparable in both versions.
After several iterations where we increased the management database size and compared CPU utilization on each run, we found the Pentium Pro 200 to
be a good match for the SPARC 10. In fact, as the database grew, we found that the Pentium Pro kept slightly ahead of the dual-processor SPARCserver by generating approximately 5 percent to 10 percent less CPU utilization in each test. In every iteration, interactive performance remained comparable.
To push the two platforms to the limit, we unleashed a complete network discovery, adding all possible nodes, using a range-test discovery on an entire Class B IP network. The resulting database of 19,000-plus models represented a larger-than-recommended database for a single server. Both servers surprised us by keeping up, although running the SpectroGRAPH interface on the same machine did stress both platforms beyond design limitations.
After loading the more than 19,000-model databases, we found that the SpectroSERVERs on both platforms intermittently refused connections--though to be fair, Cabletron doesn't recommend running both the server and SpectroGRAPH on the same machine. When the interfaces ra
n from another machine, we found performance noticeably sluggish on both the Solaris and NT platforms, and CPU utilization on the NT side remained, on average, lower than that of the Solaris server.
We found that NT performed well alongside the Sun SPARCserver. In fact, that slightly sluggish interf
ace we initially noticed actually masked a well-performing back-end server. We found NT to be a surprisingly adequate platform for network management, running Cabletron's SPECTRUM.
Bone of Contention
With its low cost for hardware and software, ease of implementation and administration, and availability of applications, NT is fast-becoming the platform of choice in small-to-medium-scale application server environments. Similarly, environments where performance or scalability is the primary concern will most likely continue to support Unix. However, concern for cost and scale lies the true battleground between NT and Unix.
Though Windows NT is available on Digital's Alpha architecture, the Int
el processor is definitely at the heart of NT's recent growth--Microsoft estimates that more than 90 percent of NT installations are x86-based. At press time, the most powerful Intel processor architecture was a six-processor 200-MHz Pentium Pro from Advanced Logic Research (ALR), with eight-processor machines soon to arrive. While seemingly limited vis-ý-vis highly parallel Unix systems, the performance of the Pentium Pro processor drives a stake into the argument against PCs as high-end servers. In our labs, single-processor Pentium Pro servers are roughly equivalent to RISC platforms like Sun's SPARCserver with two processors. NT's scalability currently is limited only by the processing power of the Pentium Pro; although the Alpha platform is faster, application support pales compared with that for the x86 platform.
Remember, though, that scalability doesn't necessarily equal processing power. The latest buzzword in server performance is clustering. Instead of relying on faster processors and buses to
deliver ever-faster servers to handle ever-increasing loads, clustering technologies promise to spread the load among a farm of small, inexpensive servers.
Akin to RAID's effect on disk subsystems, server clustering promises to use multiple inexpensive servers that work together to spread load and increase
reliability so that when a server fails, its peers will absorb its workload. However, since clustering technologies are still in their early stages, clustering won't hit the mainstream NT market until Microsoft releases Wolfpack, a standardized clustering API.
Wolfpack will deliver the operating system-level middleware and management tools that will enable peer servers to collaborate to form a cluster. The first step--high-availability clustering whereby a second server can automatically fail-over and take the place of a failed server--is slated to be available this year. However, if you're waiting to solve NT scalability problems with Wolfpack, you'll be waiting into 1998. And by then, the NT ver
sus Unix debate may be playing dead.
Dan Backman can be reached at dbackman@nwc.com.
|
REPORTS
ANALYTICS
Follow 
