From ted@squall.cis.ufl.edu Sun Oct 31 18:36:13 1993
Received: from squall.cis.ufl.edu by SHASHA.CS.NYU.EDU (5.61/1.34)
	id AA20610; Sun, 31 Oct 93 18:36:07 -0500
Received:  by squall.cis.ufl.edu (5.61ufl/4.12)
	id AA00561; Sun, 31 Oct 93 18:34:12 -0500
Date: Sun, 31 Oct 93 18:34:12 -0500
From: "Ted Johnson" <ted@squall.cis.ufl.edu>
Message-Id: <9310312334.AA00561@squall.cis.ufl.edu>
To: shasha@SHASHA.CS.NYU.EDU
Status: RO

Dennis,
I put together a set of figures for the picture.
Here's what they mean:

1) This compares A1in-only and A1out-only
2) investigate the effect of the size of A1in on hit rate
3) "   "
4) compare all algorithm with zipf input
5) "    "
6) Show why 2Q, LRU/2 work (hit rates highest on hot items)
7) I wanted to investiate responsiveness. I ran the algorithms for
   1,000,000 zipf references, then made a random permutation of the
   probabilities assigned to the items (so its a radical change of
   locality). This chart shows hit rates as a function of time after 
   the permutation.
8) "   "
9) The next charts examine algorithm performance when the input contains
   scans.  1/3 of references are scans, 2/3 are zipf.  this particular
   chart investigates the sensitivity to the size of A1in (not much).
10) compare algorithm performance
11) "  "
12) look at degredation of hit rate as compared to zipf-only
13) "  "
14) run the algorithms on the windows traces.  I modified LRU/2 to use
     a FIFO buffer, as we discussed, and used the FIFO size that gave
     the best possible hit rate (there was an improvement)
15) the two-pool experiment.

I make my plotting package produce postscript for a HP IIID.
If you still can't print, I'll try using a different font.
	Ted

