Yes, this is a pretty real problem with the new page cache: the
"allocation block" for the cache is larger than before, and as such the
page cache can be slower than it used to be under 1.2.x.
I'm afraid that is inevitable - there is a tradeoff here, and overall the
page cache is so much cleaner than the old buffer cache that I'm not
looking back (especially as most of my machines have indecent amounts of
RAM in them considering how much I actually would need).
This is a real problem for development, actually - most of the people
who test out the development kernels have reasonably high-end hardware,
if only because on low-end hardware it can take longer to compile a
kernel than it takes for me to release a new one ;-). Anyway, that kind
of leads to the kernels getting much more testing on high-end machines. I
hope we're still usable on a 4MB machine ;-/
Linus