>Since paging behavior is by inherently unpredictable, readahead
>is of quite limited use for paging.  To do readahead you need to
>evict pages you know were used in favor of pages you don't know
>will ever be needed.

OTOH, it does not pay to use an extremely small pagesize.  This
is reflected in several ports which combine physical pages to
create a logical page for paging ops, etc.

Intuitively, it would seem that the logical pagesize should
correlate to the square root of the main-memory size, eg.
4096 4KB pages on a 16MB box, 32768 32KB pages on 1GB.  In this
way things like work factors and queue lengths tend to scale
linearly with memory size.

I doubt that you would want to create a bigger logical pagesize
by reworking the pmap code -- which is machine dependent anyhow.
But you might be able to come up with a scheme that aggregates
paging activity on some-power-of-two pages.