Path: utzoo!utgpu!jarvis.csri.toronto.edu!rutgers!tut.cis.ohio-state.edu!ucbvax!agate!darkstar!june.cs.washington.edu
From: pardo@june.cs.washington.edu (David Keppel)
Newsgroups: comp.os.research
Subject: Re: compress/decompress paging?  prepaging?
Message-ID: <280@darkstar.ucsc.edu>
Date: 14 Jun 89 17:00:14 GMT
Sender: usenet@darkstar.ucsc.edu
Organization: U of Washington, Computer Science, Seattle
Lines: 42
Approved: comp-os-research@jupiter.ucsc.edu


wilson@carcoar.stanford.edu (Paul Wilson) writes:
>[I'm looking for information on paging with compression.]
>[Goal: use compression for RAM caching of pages]

Our moderator writes:
>[It has to be FAST]
>[Pages going to 2ndary store aren't fixed-size]

Here's my guess:

* Pages are compressed in-core.  Unless you know a lot about the data
  you're compressing, a fast algorithm probably won't buy much more
  than 50%.  Probably separate compression algorithms are in order for
  the code space and the data space(s).

* The pages are stored in-core in adjacent memory.  (This is what you
  would assume:  If page 1 is 100 bytes compressed and stored at
  location 2000, then page 2 is stored starting at location 2100).

* You'll want to store the compressed data pages on full-page
  allocations.  When you want page _x_, you don't want to have to
  figure out which sectors it lives on, you just want to go get it.
  Indeed, if bandwidth is high, you might want to store them on the
  disk in *uncompressed* form.  If you need it, you're going to need
  it uncompressed (unless you are doing clever prefetching).

* You may have some problems on some architectures.  The VAX, for
  instance, stores 4 bytes for each page.  The bytes include some
  protection info, so there are less than 32 address bits per page.
  Those 32 bits are used normally for either (a) the physical page
  frame number (22 bits) or for a disk block number (?? bits, < 32).
  If the pages are compressed, then they will live at odd addresses
  and will most generally require 30+ bits for each compressed page
  address.

Hope this is useful.

	;-D on  ( A page from the history books )  Pardo
-- 
		    pardo@cs.washington.edu
    {rutgers,cornell,ucsd,ubc-cs,tektronix}!uw-beaver!june!pardo