A Strategy Selection Framework for Adaptive Prefetching in Visual Exploration

1
A Strategy Selection Framework for Adaptive
Prefetching in Visual Exploration

• Punit R. Doshi, Geraldine E. Rosario, Elke A.
  Rundensteiner,
• and Matthew O. Ward
• Computer Science Department
• Worcester Polytechnic Institute
• Supported by NSF grant IIS-0119276.
• Presented at SSDBM2003, July 10, 2003.

2
Motivation

• Why visually explore data?
• Ever increasing data set sizes make data
  exploration infeasible
• Possible solution Interactive Data Visualization
  -- humans can detect certain patterns better and
  faster than data mining tools
• Why cache and prefetch?
• Interactive visualization tools do not scale
  well, yet we need real-time response

3
Example Visual Exploration Tool XmdvTool
Data Hierarchy
4
Example Visual Exploration Tool XmdvTool
Drill Down
Structure-Based Brush1
Parallel Coordinates (Linked with Brush1)
Roll-Up
Structure-Based Brush2
Parallel Coordinates (Linked with Brush2)
5
Characteristics of a Visualization Environment
Exploited for Prefetching
Move up/down

• Locality of exploration
• Contiguity of user movements
• Idle time due to user viewing display

Move left/right
6
Overview of Prefetching

• Locality of exploration
• Contiguity of user movements
• Idle time due to user viewing display

Users next request can be predicted with high
accuracy
Time to prefetch
Fetching
New user query
Idle time
Cache
DB
Prefetching
7
Static Prefetching Strategies
Random Strategy
Direction Strategy
8
Drawbacks of Static Prefetching

• Lacks a feedback mechanism
• Different users have different exploration
  patterns
• A users pattern may be changing within same
  session

• ? Generates predictions independent of past
  performance.

? No single strategy will work best for all users.
? A single strategy may not be sufficient within
one user session.
This calls for Adaptive Prefetching changing
prediction behavior in response to changing data
access patterns.
9
Types of Adaptive Prefetching

• Fine tuning one strategy
• Change parameter values of one strategy over time
  depending on past performance
• Strategy selection among several strategies
• Given a set of strategies, allow the choice of
  strategy to change over time within same session,
  depending on past performance

10
Strategy Selection

• Requirements for strategy selection
• Set of strategies to select from
• Performance measures
• Fitness function
• Strategy selection policy

11
Set of Strategies Performance Measures
Strategies
Performance measures
Required by user
Predicted by prefetcher
12
Fitness Function
Fitness function

• Other fitness functions
• global average misclass. cost
• local average response time
• global average response time

13
Fitness Function Definitions

• Global Average

Local Average (using exponential smoothing)
14
Strategy Selection Policy

• Strategy selection policies
• Best
• Proportionate

15
Performance Evaluation

• Setup
• XmdvTool as testbed
• 14 real user traces analyzed
• User traces were analyzed for
• Tendency to move in the same direction
• Frequency of movement
• Size of sample focused on
• 3 user types random-starers, indeterminates,
  directional-movers
• We will show
• Detailed analysis and results for 2 user traces
• Summary results for all user types

16
Directional User Navigation Patterns Over Time

• Ave 73
• directional
• Ave 70
• queries/min
• Navigation
• pattern
• changes
• over time

17
Directional User Navigation Patterns Over Time
Move up or down then move left to right to left
18
Directional User Directional prefetcher is best
Selection matched more directional navigation pat
tern. Any kind of prefetching is better than
none.
19
but SelectBest is even better
SelectBest chose Directional No-Prefetching
No-Prefetching selected when queries/min is
high dir is low.
20
Directional User Other performance measures
Misclassification cost trade-off between NP
MP. SelectBest gave low NP and high MP.
21
Directional User Other performance measures
SelectBest gave best CP response time but
this will not always be the case. Choice of
fitness function is important.
22
Indeterminate User Navigation Patterns Over Time

• Ave 50
• directional
• Ave 40
• queries/min
• Pattern
• changes
• over time
• Move left
• then perturb
• up down.
• Move right
• then perturb
• up down.

23
Indeterminate User SelectBest is better
SelectBest chose Random No-Prefetching
No-Prefetching selected when queries/min is
high dir is low.
24
Summary Across All User Types
Experiments repeated 3x and averaged. Reduced
prediction error for random-starters and
directional-movers. No improvement in response
time.
25
Related Work

• Adaptive Prefetching
• Strategy Refinement - Davidson98, Tcheun97,
  Curewitz93, Kroeger96, Palpanas99
• Learning - Agrawal95, Swaminathan00
• Adaptation Concepts Mitchell99, Waldspurger94,
  Avnur00
• Performance Measures Joseph97,Weiss25,
  Mitchell99
• Database support for Interactive Applications
  Stolte02, Tioga96

26
Observations

• Prefetching is better than no prefetching
• Different users have different navigation
  patterns, same user has varying navigation
  patterns within same session
• No single prefetcher works best in all cases
• Strategy selection allows prefetcher to adapt
• Performance of strategy selection depends on
  fitness function being optimized

27
Contributions

• The first to study adaptive prefetching in the
  context of visual data exploration
• A proposed framework for adaptive prefetching via
  strategy selection, as opposed to common approach
  of strategy refinement
• Empirical results showing benefits of strategy
  selection over a wide range of user navigation
  traces

28
Thats all folks

• XmdvTool Homepage
• http//davis.wpi.edu/xmdv
• xmdv_at_cs.wpi.edu
• Code is free for research and education.
• Contact author rundenst_at_cs.wpi.edu