it was suggested that my previous results from "make build" weren't
very interesting since the test machine had 128MB of RAM, and that
a test with 32MB would demostrate some performance problems.
I tried this, and in doing so I noticed that I had run the previous test
with a debug kernel, so I reran the previous test with a non-debug kernel
as well.  the results were:


1.5 32MB
3126.891u 596.892s 2:00:53.25 51.3%     0+0k 151929+207357io 25155pf+0w

-current+balancing 32MB
3253.558u 723.490s 1:39:03.70 66.9%     0+0k 10687+30709io 14532pf+0w


1.5 128MB
3092.888u 560.706s 1:36:37.29 63.0%     0+0k 51347+197915io 11051pf+0w

-current+balancing 128MB
3252.077u 746.587s 1:34:11.46 70.7%     0+0k 7003+30014io 4295pf+0w


so it appears that the more flexible usage of memory with UBC allows
for much better performance that the 1.5 scheme when there's less memory
available.  it would no doubt help the 1.5 results to increase the amount
of memory assigned to the buffer cache, so if anyone wants to play
around with that I'd be interested in hearing what the best value is
for this kind of workload.  but the results with the default tuning
values are very important, since that's what many people will use.

(the kernels configs I used were just GENERIC with unused devices drivers
removed and DIAGNOSTIC turned off.  I used "options REALEXTMEM=31744" to
reduce the effective amount of RAM.  I can provide additional details of
configuration if other people try this and get vastly different results.)

-Chuck