CS848 Paper Review Form - Fall 2006

Paper Title: Improving Performance of Internet Services Through Reward-Driven 
Request Prioritization

Author(s):

1) Is the paper technically correct?
 [X] Yes
 [ ] Mostly (minor flaws, but mostly solid)
 [ ] No

2) Originality
 [ ] Very good (very novel, trailblazing work)
 [X] Good 
 [ ] Marginal (very incremental)
 [ ] Poor (little or nothing that is new)

3) Technical Depth
 [ ] Very good (comparable to best conference papers)
 [X] Good (comparable to typical conference papers)
 [ ] Marginal depth
 [ ] Little or no depth

4) Impact/Significance
 [ ] Very significant
 [X] Significant
 [ ] Marginal significance.
 [ ] Little or no significance.

5) Presentation
 [ ] Very well written
 [X] Generally well written
 [ ] Readable
 [ ] Needs considerable work
 [ ] Unacceptably bad

6) Overall Rating
 [ ] Strong accept (very high quality)
 [X] Accept (high quality - would argue for acceptance)
 [ ] Weak Accept (marginal, willing to accept but wouldn't argue for it)
 [ ] Weak Reject (marginal, probably reject)
 [ ] Reject (would argue for rejection)

7) Summary of the paper's main contribution and rationale
   for your recommendation. (1-2 paragraphs)

This paper proposes a reward-driven request prioritization (RDRP) algorithm 
that assigns higher execution priority (on web application server resources) 
to those client web sessions that are likely to bring more profit to the 
service. The main idea is to predict the future session structure of the 
client's requests based on his interactions seen so far augmented with 
aggregated information about client behaviors and give preferential services 
to clients whose sessions are predicted to have more reward. The solution is 
shown to work better, as compared to other similar techniques, under both 
server under-load and overload conditions.

The idea presented in the paper is significant as it provides a very neat 
way of distinguishing between more profitable and less profitable clients 
from a service provider's point of view thus giving it more control over 
retaining high-valued customers by ensuring high QoS to them. The work has 
been well presented with sufficient technical depth exhibiting the use of 
statistical and customer behavior modeling techniques. The paper also shows 
that per-client history-based approach matches the performance of RDRP only 
under high correlation between client's past and future behavior. Good work 
overall.

8) List 1-3 strengths of the paper.  (1-2 sentences each,
identified as S1, S2, S3.)

S1: As opposed to the static identity based prioritization techniques the 
proposed mechanism allows to assign/change priorities to client requests 
dynamically.

S2: The idea presented is novel, practical in nature and is likely to find 
many applications in the real world internet service provider scenarios.

9) List 1-3 weaknesses of the paper (1-2 sentences each,
identified as W1, W2, W3.)

W1: The paper just comes up with a ‘magic’ number for values such as timeouts 
for obtaining a thread or the size of DB connection pool without giving any 
information/guideline about how that number was chosen.

W2: All the CBMGs that the paper presents perform a transition from 
Register-> Buy Request -> Buy confirm states each with a probability of 1 
which I think is not always true. 

W3: The layout of the headings/sub-headings in the paper is not very 
convenient for the reader giving a hard time in relating each main section 
with the corresponding sub-sections. The problem is aggravated by the fact 
that almost same font size/style was used for both headings and sub-headings.


10) Detailed comments for authors.

I would just like to add a few comments.

Regarding W2 stated above. I personally have gone up to the point of putting 
items in my cart and even registering but going no further, either returning 
to search items or exiting thus my session did not result in a purchase. If 
many clients do things like I did, then RDRP will incorrectly assign higher 
priorities to such clients when used with the presented CBMGs.

Regarding the collection of client request behavior the first thing that comes 
to my mind is the similar scenario of using cookies over the past years for 
accumulating client’s web history. It is a well-known issue that various marketing 
companies exploited the use of cookies resulting in client’s getting numerous 
unwanted emails (spam) from different advertisers which ultimately lead to user 
frustration. So one thing that I am particularly interested to know is that do 
the mechanisms used in ‘Request Profiling’ for gathering accumulated client behaviors 
devised in a way that they cannot be exploited in a similar fashion ? How about 
conserving customer’s privacy? It was not clear from the paper that whether RDRP 
allows the service provider to give preference to certain clients maintaining their 
anonymity or the clients will have to reveal their identity somehow to get that benefit?