Microsoft fudges Win2K speed trials
Windows 2000 not faster than NT after all...
Microsoft has fudged the results of Windows 2000 Professional speed tests in an attempt to convince potential purchasers that the OS offers better performance than NT Workstation 4.0. In the press release announcing the shipment to manufacturing of Windows 2000 on 15 December, Microsoft said "new benchmark results illustrate the specific performance benefits that customers will receive with Windows 2000. These tests show that Windows 2000 Professional outperforms Windows 95, Windows 98 and the Windows NT Workstation 4.0 operating system running on a computer with a minimum 64MB of memory". This claim is not sustained by the data. The same press release also states that "Windows 2000 Professional is the fastest Windows client yet. Independent tests conducted by Ziff-Davis Labs and IT Week show that Windows 2000 Professional is... up to 24 per cent faster than Windows NT Workstation 4.0 in configurations with 64MB of memory or higher". Again, the "up to 24 percent faster" claim does not follow from the ZD report. Microsoft's introduction to the ZD Labs report, Microsoft Windows 2000 Professional -- System Performance Comparison, states: "In tests running the most popular business applications, Windows 2000 is... significantly faster than Windows NT 4.0 on configurations with 32MB. It is comparable to the performance of Windows NT 4.0 on configurations with 64MB and 128MB of memory, according to the November report by ZD Labs." So already we have gone from "outperforms" to "up to 24 per cent faster" to "comparable". The ZD report tells another story that Microsoft did not see fit to mention: "On average, however, we found that Windows NT 4.0 provided slightly better performance than Windows 2000 when running with 64MB and 128 MB of RAM. Although the performance difference at 64MB was small, we found that the performance difference between the two operating systems grew when we added more memory. With 128MB of RAM, Windows 2000 was three per cent slower than Windows NT 4.0." The clear extrapolation, unfortunately not tested, is that with 256MB and more (a real world probability), NT4 would be significantly faster than Windows 2000. The executive summary of the ZD report indulges in some fudged averaging of its own to create a false impression about Windows 2000 performance. It so happens that, according to the tests, Windows 2000 performs better then NT4 with 32MB, but of course in the real world, nobody would use 32MB and, as ZD points out, "Microsoft recommends at least 64MB of memory". This slightly improved Windows 2000 performance with 32MB is just enough to tip the balance and make it seem faster "on average" -- but only just (a score of 16.2 to 16.4 in ZD's Business Winstone 99 test). As ZD points out, Microsoft requested tests with 32MB (because it no doubt knew about the performance anomaly), but carefully avoided suggesting 256MB, lest the average went in favour of NT4. This is not the end of the story, however. There is a further trick disclosed in the test system specifications and identified in the appendix to the ZD report: the NT4 tests used PIO (Programmed I/O) file transfer mode, while the Windows 95, 98 and Windows 2000 tests all used DMA (Direct Memory Access). The consequence is perfectly clear: the NT4 tests were crippled to make Windows 2000 performance seem better. PIO is performed by the CPU, and consumes cycles making performance slower. It is not suited for heavy-duty work, and is normally only used for lower-performance applications and single tasking. In DMA mode, of course, transfers take place without consuming processor cycles, just requiring an instruction to start the transfer and one to stop it. Credit is due to ZD for disclosing this information, even if it failed to point out the significance. But after all, he who pays the piper calls the tune -- and NT4 dances faster than Windows 2000. ®
Sponsored: Hyper-scale data management