BENCHMARKING Windows and OS/2

December 11, 1992 - by Timothy F. Sipples

[This message is available as file winbench.txt.]

The following is a summary of results of the PC Magazine Winbench Version 2.51 benchmarking software when run under Windows 3.1 on DOS 5.02, Windows 3.1 NT (October public beta, CD-ROM), and OS/2 2.0x (November Professional Developer's Kit CD-ROM, the third public Win-OS/2 3.1 beta).

Note that Microsoft does not recommend using NT currently for performance benchmarking. Also please note that these are a first set of benchmark figures for one particular benchmark. I hope to have numbers on a mixed sample Excel macro and on a disk I/O test. However, these numbers may be instructive.

Insofar as possible an effort was made to compare apples to apples. Tests were conducted by IBM under the supervision of another customer and myself.

The hardware consisted of an IBM PS/2 Model 95 with an 80486DX 33 MHz processor, 16 MB of RAM, XGA(-1) graphics adapter, and a 320 MB, 12 ms SCSI hard drive. (Aside from the addition of a CD-ROM drive and Sound Blaster adapter, this machine was a stock Model 95, including a busmastering 32-bit SCSI cache adapter. A standard PS/2 mouse was attached and was not moved while the benchmark was running.) (Other tests were performed on an IBM PS/2 Model 57 with standard VGA for comparison. These results are outlined in brief, below.)

Note that Winbench measures graphics performance using the default weighting system from PC Magazine (and not set by IBM). As such, the hard disk details are not particularly relevant, since disk I/O throughput is not being measured. (With luck, I will try to obtain those results and release them in the near future.) Winbench runs entirely from memory, so disk I/O is not a factor. This benchmark measures pure graphics performance. However, on the slightest chance that disk I/O would skew any results, Winbench was loaded from the same location on disk for performance measurement under each operating system (environment). Also, just for the record, the hard disk was partitioned with 200 MB to OS/2 2.0x and 100 MB to Windows NT and Windows 3.1 on DOS 5.02. All partitions were FAT. To install Windows NT Beta and OS/2 2.0x on the same hard disk required zeroing out the ID of the OS/2 Boot Manager partition. Windows NT Beta includes a specific check which is deliberately designed to refuse installation if OS/2 Boot Manager is present on the hard disk. (OS/2 2.0x includes no such NT check.) Zeroing out the Boot Manager ID renders NT Beta installable while preserving OS/2 Boot Manager capabilities.

All system defaults at install time were used. (Exception: OS/2 2.0x had PRIORITY_DISK_IO set to NO, FILES=100, and PRINTMONBUF=402,0,0. These settings are disk and printing related and do not impact Winbench. Also, DOS was loaded HIGH for the OS/2 2.0x DOS sessions, but every other default was retained for the Win-OS/2 sessions. The DOS HIGH setting also should not impact Winbench. DOS HIGH was the default on the other two.) Winbench was the only task running on each system when run.

Windows 3.1 on DOS 5.02 was installed with SMARTDRV using all default install time settings. Windows NT was installed with all default settings. All used 1024x768 in 256 colors for their desktops.

[The "Optimized" OS/2 2.0 figure consists of the following changes to the session's DOS Settings: VIDEO_RETRACE_EMULATION from ON to OFF, EMS zeroed/disabled, XMS_MEMORY_LIMIT to 64K, IDLE_SENSITIVITY to 100, DDE and Clipboard changed from Public to Private, and VIDEO_8514A_XGA_IOTRAP from ON to OFF. These changes were made more to satisfy my curiosity than anything else.]

Winbench supplies a final index number called the Graphics WinMark. This figure is the one presented below.

As a baseline comparison, Windows 3.1 on DOS 5.0 on a Compaq 386/25e running with standard VGA at 640x480 in 16 colors generates 1,676,475 WinMarks. (The higher the number of WinMarks, the faster the graphics performance.)

Double trials were run in certain cases when the results seemed surprising. These double trials also give a sense of the small variance inherent in the Winbench benchmark.

Here are the results: Exact numbers are not available to me yet, but on a stock IBM PS/2 Model 57 with 16 MB of RAM and standard VGA, the relative Winmarks were approximately as follows:

Preliminary Conclusions
The major surprise to me was the sluggish graphics performance in Windows NT Beta, which is borne out by subjective perceptions as well. (I was heard to exclaim, "This is much faster" when the switch was made from Windows NT Beta to Windows 3.1 on DOS 5.02.)

At first I suspected that the peculiarity was confined to the XGA driver (perhaps a particularly bad implementation). But the VGA numbers seem to confirm that it isn't a driver issue (although it does point out that OS/2 takes good advantage of busmastered coprocessed video). (As another reference point, Windows 3.1 on DOS 5.02 generated 3,187,612 Winmarks on the same IBM Model 95 when it was dropped down to VGA 640x480 16 color mode. Thus, Windows 3.1 on DOS benefits from XGA as well.) The Model 57 (not the 486SLC model, which was not tested) uses relatively slow, vanilla VGA hardware, and yet NT Beta suffers dramatically while OS/2 holds its own with Windows 3.1 for DOS, so it seems that OS/2's good showing is not due to any peculiarity with XGA.

In any event, the clearest conclusion one can draw from these numbers, it seems, is that Windows NT Beta has a long way to go toward even coming close to the graphics performance in either OS/2 2.0x Win-OS/2 or Windows 3.1 on DOS.

One might argue that these numbers are based on a 16-bit Windows benchmark, and that NT might really shine with a 32-bit benchmark. Perhaps true, but when benchmarking one must try and relate the benchmarks back to reality. The vast installed base of PC software is 16-bit, and as users migrate to a new 32-bit operating system they will carry with them lots of 16-bit software. It will be some time, even under the rosiest scenarios, before any large portion of that software base is replaced with the 32-bit versions. And that 32-bit version could just as easily (if not more easily, since OS/2 2.0 has been a released, shrink-wrapped product for many months with hundreds of shipping 32-bit applications) be an OS/2 2.0 32-bit application as it could be a Windows NT 32-bit application.

One might also argue that running Win-OS/2 full screen is overstating the OS/2 2.0x performance in relation to NT, since NT is doing something similar to OS/2 (in running Windows 16-bit applications "seamless") and since users will often run Windows applications "seamless" under OS/2 2.0x. Perhaps, and while the seamless Win-OS/2 numbers are not presented above, at least OS/2 2.0x gives the option of providing a mode which executes existing 16-bit Windows applications as fast as possible (in a desktop which is just like Windows for DOS - or Windows NT, for that matter - namely with a full screen Windows desktop) as well as the lower performance, "seamless" mode. NT does not offer both modes of operation - there is but one desktop.

Then there is the issue of standard v. enhanced mode, and OS/2's capabilities in running applications which require WINMEM32.DLL. The benchmark numbers were provided for Windows 3.1 in Standard Mode on DOS 5.02, for comparison, so that apples can be compared to apples. (In certain areas, Windows 3.1 Standard Mode is faster than 386 Enhanced Mode, but this is not reflected in this particular benchmark.) The services provided by WINMEM32.DLL will not be available under Windows NT - such applications will not run in the future under either operating system. This consideration is important when comparing the ease with which one can migrate to a 32-bit operating system. (The version of OS/2 2.0x under consideration can run with Windows real mode and Windows 3.0 kernels, as an option, simultaneously with the Windows 3.1 kernel, to provide better backward compatibility than Windows 3.1.)

These results are preliminary and based on beta code. And the Winbench measure has its flaws. However, without dramatic performance improvements in Windows NT, OS/2 2.0 seems to offer dramatically better performance for the large library of 16-bit Windows 2.x and 3.x applications, thus protecting that substantial software investment. With XGA, OS/2 2.0x even outpaces Windows 3.1 for DOS in raw graphics performance, as measured in Winmarks. Windows NT Beta, on the other hand, when running on a speedy PS/2 Model 95, lags well behind even the baseline Compaq 386/25e in standard VGA, which is a system that is roughly half the speed of the Model 95 in almost every dimension.

Note also that these tests were conducted on a 16 MB system, which, if anything, skews the results in favor of Windows NT Beta. The results may well have been even more striking with an 8 MB system, a more typical memory configuration. Windows NT Beta suffers performance problems on "low" RAM configurations (and, in this case, 8 MB qualifies). OS/2 2.0x, on the other hand, is very comfortable with such a RAM configuration. Winbench, when running on such a "constrained" system, may well have caused NT to page, thus impacting the results.

I hope my description of the methodology used was clear, and I hope to be providing additional numbers in the near future.

T.F.S.