Power Mac G3/266

For months now, the beige Power Mac G3 has been the most popular profile on Low End Mac, indicating to us that it's the hottest low-end Mac out there. We recently rated it a Low End Mac Best Buy - but with reservations if you want to run Mac OS X.

We picked up a beige Power Mac G3/266 in January 2002 as our testbed machine for learning Mac OS X - and seeing how well (or poorly) X does on a slower, older Mac with poorly supported graphics acceleration.

Before that, we wanted to run the usual benchmarks. The CPU runs at 267.28 MHz with a 512 KB 133.68 MHz level 2 cache and a 66.82 MHz system bus. The computer has 288 MB of RAM installed. The installed hard drive is a Quantum Fireball SE 4.3A.

This is by far the most comprehensive set of benchmarks we've published yet, ranging from OS 8.1 through 9.2.2 and covering various disk cache sizes and with virtual memory on and off to see which settings provide the best performance.

Speedometer 3.06

The system was tested on 21 January 2002 using a full install of Mac OS 8.1. Computer attached to a 12" mono monitor and tested in 8-bit video mode at 640 x 480 resolution. Results are relative to a Mac SE or Classic, which rates 1.0. Numbers rounded off to one or two decimal places.

Keep in mind that Speedometer 3.06 is written for the old 68K Macs and runs in emulation on Power Macs.

The first set of numbers compares performance at different cache settings with virtual memory disabled.

cache    CPU  graphics  disk   math 
 96KB   134.8    51.4   8.87   637.9
128KB   134.8    51.4   9.14   637.9
256KB   134.8    52.7   9.14   637.9
512KB   134.8    52.1   9.14   638.5

The cache setting should have little influence on non-disk tests, which these numbers bear out. With this particular setup, cache size makes only a small difference in performance - even disk performance beyond the 128K mark. At least as far as Speedometer 3 is concerned, a 256K or 512K disk cache seems optimal under OS 8.1.

There are several claims about virtual memory. One is that setting VM to 1 MB more than physical RAM (in this case, 289 MB) provides the best balance of speed and efficient memory use. Others claim the best results come when VM is set to a multiple of 32 MB. Still others, that making VM double installed RAM is best.

The following results are for VM off, on at 289 MB, on at 320 MB, and on at 576 MB using a 256K disk cache.

test     CPU  graphics  disk   math 
 off    134.8    52.7   9.14   637.9
289MB   134.8    51.4   9.03   637.6
320MB   125.1    51.4   9.25   637.1
576MB   125.1    49.5   9.03   635.0

These figures give some credence to the argument that some VM settings are better than others. The graphics score takes a reduction with VM enabled, and the CPU score drops 7% at the 320 MB setting, while the disk score is better at 320 MB than with VM on than with VM set to 1M above RAM. Going to 576 MB VM reduced everything except the CPU score. Best overall results with Speedometer 3 and OS 8.1 are with VM disabled.

We next updated the G3/266 to Mac OS 9.0, turned off virtual memory, restored the disk cache to 256K, ran Norton Utilities and Speed Disk on the drive, and began our OS benchmarks, using various versions of OS 9.x as well as testing with VM on (at 320 MB) and with the default (8 MB!) disk cache. Here are our results:

OS/cache       CPU     gr.   disk   math    OS/cache         CPU    gr.    disk   math 
OS 8.1/256    134.8   51.4   8.87   637.9   OS 8.1/256VM    125.1   51.4   9.25   637.1
OS 9.0/256    134.8   50.8   9.25   637.1   OS 9.0/256VM    125.1   50.8   9.48   636.2
OS 9.0/def    134.8   50.7  12.04   637.1   OS 9.0/defVM    125.1   48.9  11.67   636.7
OS 9.1/256    134.8   51.4   9.19   631.0   OS 9.1/256VM    125.1   50.8   9.54   630.1
OS 9.1/def    134.8   51.4  11.76   631.5   OS 9.1/defVM    125.1   51.4  11.85   630.1
OS 9.2.2/256  125.1   51.4   9.14   633.3   OS 9.2.2/256VM  125.1   50.1   9.36   631.9
OS 9.2.2/def  134.8   50.7  11.76   631.5   OS 9.2.2/defVM  125.1   48.9  11.49   631.9

This Mac performs as well on the disk benchmark without VM on OS 9.0 as it did on 8.1 with VM on -- and sees still better disk performance when VM is enabled. The default 8 MB disk cache under OS 8.5 and later really boosts disk performance. Under 9.0, overall best disk performance comes with the default cache and VM off.

Moving from 9.0 to 9.1, we regain the graphics performance lost going from 8.1 to 9.0, although math performance drops about 1%. Upgrading to Mac OS 9.2.2 surprisingly drops the CPU score without VM to match that with VM enabled - I retested to make sure it wasn't a fluke.

Next we go full circle and test the various disk cache and VM settings under Mac OS 9.2.2. The first set of tests is run with VM off.

cache     CPU  graphics  disk   math 
128KB    134.8    50.7   9.42   633.3
256KB    125.1    51.4   9.14   633.3
512KB    134.8    51.4   9.25   631.0
default  134.8    51.4  10.40   630.7

The low CPU score at 256K seems spurious, but repeated tests confirm the result. Looking at all the results here, the edge goes to the default 8 MB disk cache for top graphics, disk, and CPU scores.

This set of test compares performance with a 256K disk cache and various virtual memory settings.

test     CPU  graphics  disk   math 
 off    125.1    51.4   9.14   633.3
289MB   134.8    50.1   9.08   632.4
320MB   125.1    50.1   9.36   631.9
576MB   134.8    50.1   8.43   632.4

All things considered, we'd give VM off a small edge here. It has the best score in two benchmarks and takes second place for disk performance. Overall, though, results are quite close except for the 8.43 disk score with VM set at 576 MB.

Based on these two sets of results, we can expect the best overall performance with VM disabled and the default cache, but which VM setting will provide the best overall results with an 8 MB disk cache?

test     CPU  graphics  disk   math 
 off    134.8    51.4  10.40   630.7
289MB   134.8    51.4  10.40   631.0
320MB   134.8    51.4  11.67   629.8
576MB   134.8    51.4  11.49   632.4

If you are going to use virtual memory under Mac OS 9.2.2, you'll get the best overall results by setting VM to the next multiple of 32 greater than the amount of physical RAM installed in your computer. Your next best bet is to set VM at twice physical RAM.

We plan to run Speedometer in classic mode once we have OS X installed to see how it compares with native OS 9.2.2 performance.

Speedometer 4.02

The system was tested on 21 January 2002 under a full install of Mac OS 8.1. Computer attached to a 12" mono monitor and tested in 8-bit video mode at 640 x 480 resolution. Results are relative to a Quadra 605, which rates 1.0. Numbers rounded off to one or two decimal places.

The first set of numbers compares performance at different cache settings. (Because Power Macs don't support 1-, 2-, or 4-bit video, Speedometer 4 was unable to test the graphics.) Virtual memory was disabled for these tests.

cache   CPU  graphics  disk    math
 96KB  21.05   n/a     3.46   746.2
128KB  21.05   n/a     3.53   745.6
256KB  21.05   n/a     3.54   745.5
512KB  21.04   n/a     3.55   745.3

The cache setting should have little influence on non-disk tests, which these numbers bear out. As above, with this particular setup, cache size makes no significant difference, except that the disk score is a bit lower with a 96K cache. As with Speedometer 3, we'd call a 256K or 512K disk cache optimal under Mac OS 8.1.

There are several claims about virtual memory. One is that setting VM to 1 MB more than physical RAM (in this case, 289 MB) provides the best balance of speed and efficient memory use. Others claim the best results come when VM is set to a multiple of 32 MB. Still others, that making VM double installed RAM is best.

The following results are for VM off, on at 289 MB, on at 320 MB, and on at 576 MB using a 256K disk cache.

test    CPU  graphics  disk    math
 off   21.05   n/a     3.54   745.5
289MB  21.03   n/a     3.57   745.0
320MB  21.04   n/a     4.72   745.5
576MB  21.02   n/a     3.56   744.6

These figures give some credence to the argument that some VM settings are better than others. While Speedometer 4 measures an insignificant difference between no VM and VM set at 289 MB, changing that setting to 320 MB dramatically boosts the disk score by one-third! However, all three tests were lower at the 576 MB setting. Because of the significantly improved disk score at 320 MB, we'd call that the best setting for OS 8.1 and Speedometer 4.

We next updated the G3/266 to Mac OS 9.0, turned off virtual memory, restored the disk cache to 256K, ran Norton Utilities and Speed Disk on the drive, and began our OS benchmarks. Here are our results:

OS/cache       CPU    disk    math   OS/cache         CPU    disk    math
OS 8.1/256    21.05   3.54   745.5   OS 8.1/256      21.04   4.72   745.5
OS 9.0/256    21.03   3.55   745.2   OS 9.0/256VM    21.00   3.71   743.7
OS 9.0/def    21.03   4.59   744.7   OS 9.0/defVM    20.99   4.56   743.2
OS 9.1/256    20.95   3.54   741.9   OS 9.1/256VM    20.91   3.77   743.2
OS 9.1/def    20.95   4.61   742.1   OS 9.1/defVM    20.92   4.57   741.2
OS 9.2.2/256  21.03   3.47   745.2   OS 9.2.2/256VM  20.99   3.60   743.7
OS 9.2.2/def  21.03   4.51   745.4   OS 9.2.2/defVM  20.99   4.49   743.3

Where OS 8.0 saw a 33% boost in hard drive performance with VM enabled, OS 9.0 saw a much smaller 4.5% improvement at the same 256K setting. Under Speedometer 4, there is virtually no difference between the disk scores when using the default 8 MB disk cache.

There is a very small drop in CPU and math performance (0.4%) going to OS 9.1; it's certainly nothing you'd ever notice.

Next we go full circle and test the various disk cache and VM settings under Mac OS 9.2.2. This group of results shows the effect of the disk cache with VM off.

cache     CPU  graphics  disk    math
128KB    21.03   n/a     3.61   744.3
256KB    21.03   n/a     3.47   745.2
512KB    21.02   n/a     3.53   744.8
default  21.04   n/a     4.09   744.5

Speedometer 4 doesn't find a significant difference in CPU or math scores whether the disk cache is big or small, so we'll give the win to the 8 MB default cache setting because of the better disk score.

The next group of benchmarks measures performance with a 256K disk cache and various VM settings.

test    CPU  graphics  disk    math
 off   21.03   n/a     3.47   745.2
289MB  21.00   n/a     3.47   743.9
320MB  20.99   n/a     3.60   743.7
576MB  20.99   n/a     4.47   743.1

Speedometer 4 finds an insignificant difference among CPU and math scores under Mac OS 9.2.2, so we'd have to give the nod to VM set at 576 MB because of the significantly better disk score.

Based on these two sets of results, we can expect the best overall performance with VM disabled and the default cache, but which VM setting will provide the best overall results with an 8 MB disk cache?

test    CPU  graphics  disk    math
 off   21.04   n/a     4.09   744.5
289MB  20.99   n/a     4.08   743.3
320MB  21.00   n/a     4.54   743.4
576MB  20.99   n/a     4.49   743.2

If you are going to use virtual memory under Mac OS 9.2.2, your optimal setting will be the next multiple of 32 greater than the amount of physical RAM installed, although that barely edges out double physical RAM as far as Speedometer 4 is concerned.

We plan to run Speedometer in classic mode once we have OS X installed to see how it compares with native OS 9.2.2 performance.

MacBench 5

The system was tested on 21 January 2002 using a normal installation of Mac OS 8.1. Computer attached to a 12" mono monitor and tested in 8-bit video mode at 640 x 480 resolution. The disk cache was set to 256 KB for all tests. Results are relative to a beige Power Mac G3/300, which rates 1000.

Virtual memory was turned off. Graphics tests could not be run since the monitor does not support 800 x 600.

test       CPU    math    disk
 96KB      829     893     387
128KB      831     893     646
256KB      833     893     646
512KB      829     893     731
  1MB      831     893     731

Unlike the Speedometer tests, where a 128K cache was just 2-3% faster than a 96K cache, that seemingly small change made a 67% improvement under MacBench 5. Growing the cache beyond 256K also improved performance under MacBench, which uses a larger test file than either version of Speedometer. For MacBench 5 and Mac OS 8.1, a 512K or 1 MB disk cache seems to provide the best overall results.

There are several claims about virtual memory. One is that setting VM to 1 MB more than physical RAM (in this case, 289 MB) provides the best balance of speed and efficient memory use. Others claim the best results come when VM is set to a multiple of 32 MB. Still others, that making VM double installed RAM is best.

The following results are for VM off, on at 289 MB, on at 320 MB, and on at 576 MB using a 256K disk cache.

test       CPU    math    disk
 off       833     893     646
289MB      777     890     624
320MB      777     890     619
576MB      781     892     625

MacBench 5 finds a measurable (7%) reduction in CPU performance with VM enabled in OS 8.1, along with a small (3%) decrease in drive performance, and no significant change in math performance (well under 1%). Although some claim disk performance improves when VM is used, MacBench 5 belies that claim. Among these settings, running with VM off seems optimal for MacBench 5 and Mac OS 8.1.

We next updated the G3/266 to Mac OS 9.0, turned off virtual memory, restored the disk cache to 256K, ran Norton Utilities and Speed Disk on the drive, and began our OS benchmarks. Here are our results:

OS/cache      CPU   math   disk    OS/cache        CPU   math   disk
OS 8.1/256    833    893    646    OS 8.1/256VM    777    890    619
OS 9.0/256    831    890    630    OS 9.0/256VM    805    889    630
OS 9.0/def    830    890    990    OS 9.0/defVM    801    890    995
OS 9.1/256    820    886    630    OS 9.1/256VM    799    886    630
OS 9.1/def    817    886    999    OS 9.1/defVM    787    885    995
OS 9.2.2/256  828    890    622    OS 9.2.2/256VM  804    884    603
OS 9.2.2/def  830    890    984    OS 9.2.2/defVM  804    884   1000

Perhaps the biggest improvement in OS 9.0 is the 3% higher CPU score when VM is enabled, showing an improvement in the underlying VM mechanism. Curiously, we see a 3-4% drop in the CPU benchmark going from 9.0 to 9.1 but almost no change (under 0.5%) in the math score.

Next we go full circle and test the various disk cache and VM settings under Mac OS 9.2.2. The following tests were run with VM off and various disk cache sizes.

test       CPU    math    disk
128KB      829     889     549
256KB      833     893     646
512KB      831     891     706
  1MB      830     890     826
default    828     890     946

As before, the best CPU score was obtained with a 256K disk cache, although the range of CPU scores is very tight (about 0.6%), as is the range of math scores. All things considered, the best overall performance goes to the setting with the best disk score.

This set of results was obtained with a 256K disk cache and various VM settings.

test       CPU    math    disk
 off       833     893     646
289MB      803     884     598
320MB      777     890     619
576MB      801     884     625

Looking over these results, we find that MacBench 5.0 has the best overall scores with VM disabled.

Based on these two sets of results, we can expect the best overall performance with VM disabled and the default cache, but which VM setting will provide the best overall results with an 8 MB disk cache?

test       CPU    math    disk
 off       828     890     946
289MB      800     884     946
320MB      803     884     972
576MB      800     884     977

If you are going to use virtual memory under Mac OS 9.2.2, these tests indicate that you'll achieve the best results by setting VM at the next multiple of 32 higher than the amount of physical RAM installed in your Mac or double the amount of physical RAM installed - it's pretty much a toss-up this time around.

We plan to run MacBench 5.0 in classic mode once we have OS X installed to see how it compares with native OS 9.2.2 performance.

Conclusion

Each benchmark works differently, but these tests indicate that the disk cache should be set to at least 128K for decent results and 512K for better ones. As for virtual memory, the results are a mixed bag. We have turned conventional wisdom on its head - setting VM to 1M over installed RAM and to twice installed RAM are both generally inferior to using a VM setting that is the next multiple of 32 higher than your installed RAM under Mac OS 8.1.

Looking at all the results under OS 9.2.2, using the "next multiple of 32" provides the best overall VM performance, followed closely by setting VM at twice physical RAM. The old advice to set VM to 1 MB beyond physical RAM only applies if scratch space on your hard drive is extremely limited. Whenever possible, set VM to a multiple of 32 MB.

In the end, if you can run with virtual memory disabled, you may lose some efficiency in the use of available RAM but will more than make up for it with the best performance your Mac can manage. Buying enough memory so you can avoid using VM is the best way to squeeze the most performance out of your Mac.