Bear in mind that these graphs are not so interesting as a test of *disk*
speed, because what you are almost certainly seeing in your benchmarking is
the effect of the system cache.  The system cache behavior can be notoriously
unpredictable for stuff like this, esp. with UBC.

In fact, if the benchmark program is not carefully written, you may also
be seeing the effect of library (libc/stdio) buffering.

The actual transfer speeds (0-4000 Mbits - red zone) is actually pretty poor
for a fast SCSI drive.  I get closer to 2 MBytes/sec on my Power center 132
with a 300Mhz G3 in it when moving data to/from the other macs with netatalk.

You might try running to the raw device, if you aren't already.  That
should be a little more predictable.

???

-dgl-

>I produced some nice looking graphs of my hard drive speed using iozone, a 
>free opensource disk io benchmarking utility at http://www.iozone.org/.
>
>The graphs can be found at
>http://morning.ecet.vtc.edu/~jlaughli/powermac/
>I should have noted on the graph that all sizes are in kilobytes, while 
>speed measurement is in kilobits per second.
>
>For both read and write operations, it is clear to see how the disk cache 
>helps speed up transfers up until the size of the transfers exceed the size 
>of the disk cache, and then the speed is bottlenecked by the actual speed 
>of the scsi bus.
>
>The read graph has a couple of "interesting" features, though. Note the 
>holes that occur around 8 & 16k files with 1k record sizes, and the less 
>prominent holes at 2k record sizes with 32 and 512k files, and also a 
>smaller hole at 8k files x 4k records. Can these be explained? They are not 
>present in the re-read graph, which shows a much more uniform slope.
>
>The write graph isn't all that interesting, all though that little notch 
>and that little hole intrigue me. I probably should have used a different 
>scale to show more detail, but I wanted to use the same scale as on the 
>other tests.
>
>The re-write graph is perhaps the most interesting of all. What the heck is 
>going on there? I should really run the tests 10 times or so and average 
>the results, it's possible cron decided it was going to execute a job or 
>something just then.
>
>I would like to display the graphs with a finer resolution in the Z axis, 
>perhaps even a continuous gradient. I can't quite figure out how to get 
>excel to do something like that though. If anyone has tips for that, they'd 
>be appreciated.
>
>-Jeff