use of mmaped files

[Bud Davis]

gfortran/gcc  experts,

i need some help / ideas understanding the details of how this
mechanization is supposed to work.  


#1.  Don't we want to mmap the entire file when it is opened ?  Isn't
the idea that the OS can handle the low level details when the memory is
accessed a lot better than ordinary mortals code?  If we do mmap the
whole file, then writing a 2GB array would need 4GB of memory ? (2GB for
the array, 2GB for the mmap)

#2.  if we don't mmap the whole thing, one 8K page at a time seems way
to small.  how do we pick a size, or determine it dynamically ?  If a
size is chosen, how do we keep the edge cases from having horrible
performance ? meaning the program just happens to have an access pattern
that forces chunks to be unmapped and then re-mapped.

all ideas and corrections of my misconceptions are appreciated.

--bud