How memory affects system performance

If you're going to roll out or upgrade your office's (or enterprise's) Windows client systems, you probably want to know whether Windows NT or Windows 95 is faster. Knowing which OS is faster and under what circumstances can determine which one you deploy. The difference in cost is also a factor.

Design Factors
Win95 was designed as a replacement for Windows 3.x, so it runs most old DOS and 16-bit Windows applications and drivers. Win95 also runs all the new 32-bit applications, even some that were originally designed for NT. Win95 has sophisticated features such as pre-emptive multitasking and built-in multimedia support. It can run in as little as 4MB of memory on a 386-compatible processor.

NT was designed from the ground up to run business applications. Its features include true 32-bit application support, true pre-emptive multitasking and multithreading capabilities, security, fault tolerance, reliability, a secure file system (NTFS), and support for RISC-based and multiprocessor systems.

NT requires greater system resources, such as a 486-compatible processor, 16MB of memory, and 120MB of hard-disk real estate; has no backward-compatibility with DOS and Windows 3.x 16-bit device drivers; and offers limited support for older DOS and Windows 3.x applications, although it will run the majority of them.

The Test
I tested Win95, NT Workstation 3.51, and NT Workstation 4.0 on a Pentium Pro system, an HP 200MHz Vectra XA with 256KB cache. (The system came preloaded with Extended Data Output--EDO--RAM, which I tested for speed as you can see in "EDO RAM ," below.) I used our Windows NT Magazine Lab benchmark package, which we designed to execute a script over Adobe Photoshop and Elastic Reality. I ran this benchmark on the Vectra XA system with 16MB of memory and was surprised to find that Win95 was quicker than NT 3.51 and 4.0 by about 30 percent.

About this time, I received the BABCo SYSmark32 benchmark, which measures the time that real-world applications take to run predefined scripts. It is much like our benchmark but uses Adobe PageMaker, CorelDRAW 6.0, Lotus Freelance 96, Lotus WordPro 96, Microsoft Excel 7.0, Microsoft PowerPoint 7.0, Microsoft Word 7.0, and Paradox 7.0 as its test applications.

Graph 1 shows that Win95 is about 18 percent faster than NT 3.51 and, interestingly, 20 percent faster than NT 4.0 with the same system configuration. However, when I looked at the results for specific applications, I noticed that on NT 4.0, the Adobe PageMaker, CorelDRAW 6.0, and Microsoft Word 7.0 scripts ran as fast as or faster than on Win95, and the Adobe PageMaker and Paradox 7.0 scripts ran faster on NT 3.51 than either Win95 or NT 4.0.

Wondering about the difference in performance, I added 16MB of memory to the Vectra XA, for a total of 32MB, and ran the tests with Win95. I had heard that performance on Win95 was supposed to stay the same or decrease with more than 16MB of memory. So, I was surprised when it ran the scripts about 15 percent faster, as Graph 2 shows.

I executed the same test with NT 4.0 and was amazed to see that it ran more than 50 percent faster than before. This result was a 28 percent improvement over Win95 at this configuration. I ran the same test with NT 3.51, and the results were not as impressive, although I did see a gain of more than 16 percent.

With such a performance gain, I wondered whether doubling the memory to 64MB would double the performance again. I put 64MB of memory into the system and ran the tests again. The results in Graph 3 show no significant performance improvement, although NT 4.0 was the clear leader with a gain of 7 percent from the previous test. NT 3.51 was 14 percent faster, and Win95 gained only 4.5 percent.

Overall Comparisons
Seeing results that you can easily use to improve a slow system's performance is nice. The real world is not so nice and neat: You probably run multiple applications, device drivers, OS services, and System Tray applications at the same time. The BAPCo SYSmark32 benchmarking software runs a script over a single application, timing how long executing and then rebooting the system takes after each run. This approach is useful to test how well a specific application will run on a given system, but may not be a realistic test of the system.

Both NT and Win95 outperform each other, depending on how much memory you install, as Graph 4 shows. If you have fewer than 32MB of memory and run NT 4.0 or are considering an upgrade, I recommend installing at least 32MB of memory. If you run Win95 and have more than 32MB of memory, NT 4.0 can give your system a turbo charge.