When reading from many possibly large files in a fashion similar to random
access, it is usually the fastest and most efficient to use memory maps.
Although memory maps have many advantages, they represent a very limited
system resource as every map uses one file descriptor, whose amount is
limited per process. On 32 bit systems, the amount of memory you can have
mapped at a time is naturally limited to theoretical 4GB of memory, which
may not be enough for some applications.

The documentation can be found here: http://packages.python.org/smmap