Stories By Fabian Pascal 32-Bit Computing Performance Tweaks for the Rest of Us Part I: BIOS, Operating System and Processor Subsystem Optimization The Communications Subsystem DSL vs. Dialup Best of Breed System -- 2nd Generation: Processor, Graphics and Disk Subsystems Intel "Compatibility" Under Windows NT Hard to Keep Your Cool Disk Subsystem PC Advisor Column (www.dsi2000.com) The Graphics Subsystem The Processor Subsystem Selecting a Slot 1 System board On Chipsets and Buses   Also of Interest: Transferring Files With Flipdisk By Ed Kahn Wireless PC By Woody Liswood Linux: Today's Curiosity Or Tomorrow's Workhorse? By Kevin Savetz and Neil Randall Open Source Software: More Than An Emerging Alternative By Michael Robin

|Feedback To Author | Feedback To Webmaster Copyright © 1999 by Fabian Pascal. All Rights Reserved 32 Bit Computing By Fabian Pascal Performance Tweaks For the Rest of Us -- Part 2: Video and Disk Subsystems; CPU Overclocking  Updates Recommended System l Dropped SyQuest SyJet 1.5GB SCSI removable drive; currently evaluating Iomega's SCSI Jaz 2GB drive as replacement; lReplaced Plextor's ultraSCSI 32X CD-ROM drive with Pioneer's 303S 6X ultraSCSI DVD-ROM drive. Recommended Software lNo changes Benchmarks lAdded: Sonera's DisplayMate display evaluation utility. In the first part of this two-part series I showed that tweaking the BIOS, NT registry and chipset transfer rates has a no noticeable effect on the performance of business applications as measured by the Winstone 98 and 99 benchmarks. The BIOS tweaks account for all of the improvements of 5% for Winstone 98 (ranging from 3% for publishing applications to 5.5% for web browsers) and 3.5% for Winstone 99, which measures multitasked applications. Since BIOS tweaks cost nothing, however, there is no reason why you should not apply them. To recall, the criteria for reasonable tweaks are (a) require no technical expertise (b) easy and not time consuming (c) free or inexpensive (d) no disassembly or modifications of hardware (e) little or no damage risk. Are there other such tweaks that improve business applications performance on the recommended system? Note: The BIOS-tweaked performance serves as the base for the tests in this article. Swapfile Placement Because, as I showed in another disk subsystem article, the transfer rate decreases across the disk from the outside in, it is claimed that "... putting the swap file as the first file on a drive is because the drive can load things faster from the front of the drive." And according to Microsoft "... placing the pagefile on the boot partition does not optimize performance, because Windows NT has to perform disk I/O on both the system directory and the pagefile ... place the pagefile on a different partition and different physical hard disk drive so that Windows NT can handle multiple I/O requests in a quicker fashion ... The best option is to create the pagefile ... in its own partition, with no data or operating system-specific files." On the other hand, others argue that "Evidently, creating a permanent, contiguous swapfile that's just nice for your system and moving it to the outer tracks of the hard disk is the best way to optimize your swapfile ... Moving the swapfile to an alternate hard disk is nice from a theoretical point of view, but it is actually very hard to implement effectively. Needless to say, moving the swapfile to a different partition is both a waste of time and counterproductive." There seems to be agreement, thus, that placing a sufficiently large swap file in its own, first partition on a drive will maximize performance. Note very carefully, though, that the effectiveness of this technique depends on the amount of swapping to disk that occurs which is, in turn, a function of the available memory and the applications being run. Furthermore, the magnitude of the effect will also be affected by whether NT will need to expand the initial swap file and how much. If you have 256MB of RAM, and run only one or two applications that use small data files, there will be no swapping and if the initial size of the swap file is large enough, it won't require expansion. the placement may have no effect on performance. (To determine if swapping occurs and how much, you need to configure NT's Performance Monitor to record the swap file size in a log file; there are instrucions to that effect in the Microsoft Knowledge base). According to Ziff-Davis. "Business Winstone 99 scores increased under NT when we increased RAM above 64MB. There was little change in scores for Winstone 98 above 64MB." This suggests that the 99 version requires more than 64MB of memory -- the base memory for the benchmark -- and the 98 version less; but this does not tell us how big the swapfile must be to avoid expansion. To gauge the effect of swapfile placement on business application performance, I ran the Winstone benchmarks on the recommended system equipped with 64MB of RAM and configured a 25MB swapfile as follows: 1. in the same partition (H:) with NT 2. in its own partition (I:) at the end of the same drive 3. in a not used partition (D:) on another drive 4. in its own, first partition (J:) of the drive on which NT and the Winstone benchmarks It's very unlikely that owners of a best-of-breed system will have less memory and not enough disk space to create a 25MB swapfile, and if the placement does not help performance under these constraints, it certainly won't help in their absence. Figure 1 shows that performance is practically the same for all four configurations, for both Winstone 98 and each of its applications and Winstone 99. Swapfile Placement    H:  I: D:   J:  Winstone 98     31.00  31.10  31.20  31.10 Browsers    3.84  3.86  3.86  3.87  Word Processing    3.18  3.20  3.21  3.21  Spreadsheet/Database    2.44  2.43  2.45  2.44  Desktop Publishing    3.73  3.75  3.74  3.73  Winstone 99    22.70  22.40  22.50  22.60 *25MB swapfile initial size 25M *Recommended system configuration, PII/300, 64MB RAM *NT 4.0, SP4 Figure 1: Effect of placement of swap file on performance Color Depth Lowering the color depth of the display from true color (32bit) to 64K colors (16bit) does, in principle, involve degradation of image quality; but it is unlikely to be noticeable with common business applications (see graphics subsystem article). Figure 2 shows that it has practically no effect on performance either. This is probably due in part to the fact that unlike resolution and refresh rate, color depth is more sensitive to the amount of video memory available on the graphics controller, than to CPU or video card speed and, like the recommended Matrox Millenium G200, video cards come currently equipped with at least 16MB (and even 32MB) of video memory. CPU Overclocking One free and easy tweak left to the business application user is to take some minimal risk and, with proper cooling and a system board like the recommended Abit BH6 that permits it, overclock the CPU. The recommended CPU is a PII/300 with Deschutes core -- basically a PII/450 sold by Intel as a PII/300 -- which can be overclocked beyond 450MHz. Unfortunately, I do not have one, so my tests were limited by the "real" PII/300 -- the Klamath (see conclusions). While it normally runs OK at 350MHz (3.5x100), its L2 cache is rated at maximum 333MHz, so it crashes once in a while and it does not complete Winstone 99. Full stability can only be reached at 336MHz (3x112).   Base 16K Colors 336 3x112  350 3.5x100  Winstone 98     32.50  32.40  34.90 35.50 Browsers    3.95  4.01  4.47  4.52  Word Processing   3.29 3.28  3.59 3.67  Spreadsheet/Database   2.66  2.67 2.74  2.77  Desktop Publishing    3.81 3.72  4.18  4.24  Winstone 99    24.00  24.90  25.40  N/A *base: BIOS-tweaked, 300x66, 1280x1024x32bit@75Hz refresh *25MB swapfile initial size 25M *Recommended system configuration, PII/300, 64MB RAM *NT 4.0, SP4 Figure 2: Effect of color depth and CPU overclocking on performance In Figure 2, the effects of the overclocking on performance are * 336: Winstone 98: 7.5%, ranging from 3% for spreadsheet/database applications to13% for web browsing; Winstone 99: 6% * 350: Winstone 98: 9%, ranging from 4% for spreadsheet/database applications to 14.5% for browsing) Note: Given the small differences between 336 and 350 overclocking, the user should avoid instability by staying at 336Mhz. Adding Up The results indicate that each of the reasonable tweaks under consideration has no significant effect on the performance of business applications as measured by Winstone. But if the BIOS and CPU overclocking effects were additive, they could have a discernible combined impact. Figure 2 shows that applying the BIOS tweaks and oveclocking the PII/300 to 336Mhz yields a 19% gain for Winstone 98 (ranging from 13% for spreadsheet/database to 26% for browsing) and 17% for Winstone 99 multitasking. Conclusions The reasonable tweaks available to users of common business applications on stand-alone systems do not seem to have a significant effect on performance They are, however, easy, free and. in the case of overclocking, carry only a minimal risk. As is always the case with benchmarks, your mileage may vary depending on the specifics of your hardware and software configuration (see benchmark limitations and illusions). It may seem from the results that a faster CPU or overclocking is the best reasonable tweak. Note, however, the gains from overclocking, if any, will depend on how processor intensive your specific applications are and may not occur if the bottleneck is elsewhere (e.g. disk; see article on disk subsystem performance, http://www.yoyow.com/fpascal/disk.html). Thus, overclocking the PII/300 Klamath by 12% (from 300 to 350MHz) and the bus by 70% (from 66 to 112MHz) increases performance gains of 13% - 26%, which won't be very noticeable. And results by others show that overclocking a PII/350 by 38% from its base (3.5x100) to 483Mhz (3.5x138), yields a Winstone gain of only 20%, which is not very noticeable either. Furthermore, any gains from a faster CPU must be assessed against cost. Indeed, the very reason the PII/450 Deschutes sold as a PII/300 Klamath is recommended is its price. All this suggest that careful selection of best-of-breed components is more or less the best that users of business applications can do to maximize performance. Skimping on hardware and then dedicating time, effort and and/or expenses to unreasonable tweaking does not make sense, except if you happen to enjoy that sort of thing.