next-gen mapping

[Michael Hohmuth]

"Adam 'WeirdArms' Wiggins" <awiggins at cse.unsw.edu.au> writes:

> 	Can anyone explain whats different/new about the planned
> next-generation mapping tree for fiasco? I'd appreciate a detailed
> explination if anyones got the time.

The new thing about the new mapping database is that its data
structures are much smaller.  It uses compact linear arrays of fixed
sizes, instead of elements linked in binary trees, to store the
mapping trees for a given physical page frame.

The size of the current mapping database implementations (in L4/x86
and Fiasco) can become a severe problem: To support a heavily-loaded
L4Linux, one eighth of the main memory has to reserved for mappings.
This also has performance implications, as our AIM multiuser benchmark
measurements show: If we steal original (monolithic) Linux one eighth
of its main memory, the application-level performance penalty between
Linux and L4Linux decreases from 3.8 percent to 2 percent (on our
hardware).

Michael
-- 
hohmuth at innocent.com, hohmuth at inf.tu-dresden.de
http://home.pages.de/~hohmuth/