lecture 25 -- page tables & page replacement algorithms

we reviewed page tables (first discussed in lect 22).

what is the biggest diff betw caching and page replacement?
- caches are done with hardware, so must be fast and simple.
- page replacement algs are done by OS with software, so may be more
complex.  but, code takes time to run, so added complexity must
translate into better page-replacement performance for it to be worth it.

page replacement strategies:

- "optimal" - theoretical best if you could predict the future.  used
by simulations to benchmark practical algs.

- not recently used - reset 'referenced' bit every clock tick (20 ms
or so).  sort pages into 4 classes at page-fault time: 0 (not ref, not
mod), (1 not ref, mod), 2 (ref, not mod) 3 (ref, mod).  from the
lowest numbered non-empty class, choose page to evict (pick at random
if there are more than one).  class 1 pages occur when a class 3 is
ref-reset at clock-tick-time.  better to evict a mod'd page than one
that's been recently read.

- FIFO.  has bug that just 'cause a page has been read in a while ago,
might still be in heavy use, so it shouldn't be evicted.

- 2nd chance.  reset ref bit and add to back of FIFO queue.
degenerates to FIFO if queue is full of ref'd pages.  pretty
reasonable approach.

- clock method.  same as FIFO, but with a simpler implementation.
practically most popular.