


VLDB-94: North America ProgramDear Author:

I am pleased to inform you that paper 169

Title: 2Q
Author(s): Theodore Johnson and Dennis Shasha

Has been accepted for the VLDB-94 Conference.
             
The North American Program Committee met on September 15, 
and selected  20 papers out of 138 considered. 
This very selective acceptance rate represents 
a recognition of the quality of your work, and 
of the VLDB-94 commitment to technical excellence.  
It is therefore important that you revise and 
improve your manuscript along the lines suggested 
by the referees, whose comments are attached.  
The deadline for returning the camera-ready version 
is May 30, and  there  is a limit of 11 pages.
You should soon receive an envelope containing
instructions and materials for the preparation 
of your camera-ready version.   
             
Looking forward to a very exciting conference 
and to your paper presentation, I take this 
opportunity to congratulate you on your success.
             
Carlo Zaniolo
NA Program Chair

                   VLDB ' 94 Conference
            North American Program Committee
                   Referee Report 

Paper Number: 169 

Title: 2Q
Author(s): Theodore Johnson and Dennis Shasha

Numeric score from 1 to 9: Originality:       6
                           Technical quality: 7
                           Relevance to VLDB: 7
                           Presentation:      7
                           OVERALL RATING:    6

Legend. 9: Outstanding.  8:  Strong Accept.  7: Accept.  6:  Weak Accept. 
5: Neutral. 4: Weak Reject.  3:Reject. 2: Strong Reject  1:  Unacceptable.  


A Short Summary of the Rationale for Your Recommendation:

This paper is essentially last year's LRU/2 algorithm "done right."  For what
it is, it seems like a very reasonable paper.  My only real problem is that
the existence of LRU/2 somewhat deflates the degree of contribution of this
paper.


Detailed Comments to Authors:

	Ted: All reasonable comments.


(1)  On p. 4, don't talk about cold pages on the "Am queue" before you explain
what the Am queue is - otherwise it's slightly confusing (i.e., undefined).

(2)  In reporting Gclock results, be sure to explain what init_count was used.
I'd actually have preferred that you picked one setting and stuck with it.

(3)  I would have liked to see you set the Kin parameter consistently from
the start, rather than setting it to one (though I understood why 1 was okay)
and then setting it differently later for the trace runs.

(4)  What do you mean on p. 8 by "LRU/2 gives priority to index pages as it
should" - i.e., how do you define "should" in such a way that gives a hit rate
of 1/2?  I didn't quite follow your thinking there.  (With 100 page slots, why
not have more of the index and less of the data, for instance?)

(5)  Varying Kin from 20% to 30% seems a little narrow for a Kin sensitivity
analysis.

(6)  You call the differences between (2Q, LRU/2) and (LRU, Gclock) in table 1
"significant".  They actually don't seem all THAT large to me.

(7)  The figure 6 results (on tags vs. pages) are not consistent with what the
text says; this is presumably just a typo.

(8)  The analytical section is a bit dense for systems-oriented readers.

(9)  Another piece of related work appeared a few years ago in the SIGMETRICS
conference; it was by John Robinson and someone who had just joined Yorktown
from Illinois, if I remember correctly, with Robinson as the second author.
This work should also be discussed, as I think it had roughly comparable goals.





                   VLDB ' 94 Conference
            North American Program Committee
                   Referee Report 

Paper Number: 169 

Title: 2Q
Author(s): Theodore Johnson and Dennis Shasha

Numeric score from 1 to 9: Originality:       7
                           Technical quality: 7
                           Relevance to VLDB: 8
                           Presentation:      7
                           OVERALL RATING:    7

Legend. 9: Outstanding.  8:  Strong Accept.  7: Accept.  6:  Weak Accept. 
5: Neutral. 4: Weak Reject.  3:Reject. 2: Strong Reject  1:  Unacceptable.  


A Short Summary of the Rationale for Your Recommendation:

Discusses a simple and effective buffering policy,
supported by many experiments on real and synthetic data.


Detailed Comments to Authors:

The paper seems to make a solid contribution to buffering algorithms, building
on top of the LRU/2 idea.  The proposed algorithm is fast, simple, and more
effective than LRU (and LRU/2).
The authors fine-tuned the parameters of their model,
using real and synthetic databases.

Points for minor improvement:
  1) Can the 2Q method handle "looping references?" I guess not!
  2) It might be interesting to measure the CPU time for LRU/2 and 2Q

Ted: this is only one to discuss.
  3) Define the Zipf distribution, and give a citation to Zipf
     [George Kingsley Zipf "Human Behavior and Principle of Least Effort",
      Addison Wesley 1949]




                   VLDB ' 94 Conference
            North American Program Committee
                   Referee Report 

Paper Number: 169 

Title: 2Q
Author(s): Theodore Johnson and Dennis Shasha

Numeric score from 1 to 9: Originality:       4
                           Technical quality: 5
                           Relevance to VLDB: 5
                           Presentation:      7
                           OVERALL RATING:    5

Legend. 9: Outstanding.  8:  Strong Accept.  7: Accept.  6:  Weak Accept. 
5: Neutral. 4: Weak Reject.  3:Reject. 2: Strong Reject  1:  Unacceptable.  


A Short Summary of the Rationale for Your Recommendation:

The paper describes a 2Q algorithm as an implementation alternative
to LRU/2.  As such, its contributions are a marginal improvement 
to O'Neil and Weikum's paper that appeared in Sigmod 1993.


Detailed Comments to Authors:


Both the analysis and the algorithm for 2Q are fine.  I 
greatly appreciated your note to the VLDB reviewer (I hope 
your style becomes a standard in the future).  I did not see 
enough contributions in the 2Q algorithm (as compared to
either LRU/2 or Gclock) to recommend this paper.  In
particular, having some experience with simulation models,
I have serious reservations calling a 3% to 12% improvement
observed by 2Q as compared to Gclock a significant improvement
(Table 1 described in Section 3.2).





	Ted: Let's add the table comparing this with the Tandem
	approach. It would be good to include the Bates tapes too.
	Does this seem reasonable?

	Dennis
	P.S. When the time approaches, I'll send you my talk slides.
	Let's try to work out the talk together.
