Load%20Analysis%20and%20Prediction%20for%20Responsive%20Interactive%20Applications

1
Load Analysis and Predictionfor Responsive
Interactive Applications

• Peter A. Dinda
• David R. OHallaron
• Carnegie Mellon University

2
Overview
Load Analysis
Time Series Modelling
Measurement
History-based Load Prediction
Communication
Computation
Execution Time Predicition
Remote Execution
Best Effort Real-time
Responsive Interactive Applications (eg, BBN
OpenMap)
3
OpenMap (BBN)
Move North
Integrator
New map data
Choice of Host
Bounded Response Time
Replicated Specialists
Terrain
Terrain
Terrain
4
Context
Advanced Mobility Platform
Logistics Anchor Desk
METOC Anchor Desk
Other Applications
JTF Planner
TRACE2ES
Applications
...
Frameworks
OpenMap (BBN)
QuO (BBN)
Adaptation
Load Prediction (CMU)
Prediction
Remos (CMU)
Measurement
Distributed system
Distributed system
5
Load Analysis and Prediction

• Goal accurate short term predictions
• Few seconds for non-stale data
• Evaluation/comparison issues
• Load generation vs. Load prediction
• Have to discover which properties are important
• Performance measure
• Mean squared prediction error
• Lack of lower bound to compare against
• Simple, reasonable algorithm for comparison

6
Load Trace Analysis

• Digital Unix one minute load average
• Four classes of hosts (38 machines)
• 1 Hz sample rate, gtone week traces, two sets at
  different times of the year
• Analysis results to appear in LCR98
• Load is self-similar
• Load exhibits epochal behavior

7
Self-similarity Statistics
8
Why is Self-Similarity Important?

• Complex structure
• Not completely random, nor independent
• Short range dependence
• Excellent for history-based prediction
• Long range dependence
• Possibly a problem
• Modeling Implications
• Suggests models
• ARFIMA, FGN, TAR

9
Load Exhibits Epochal Behavior
10
Epoch Length Statistics
11
Why is Epochal Behavior Important?

• Complex structure
• Non-stationary
• Modeling Implications
• Suggests models
• ARIMA, ARFIMA, etc.
• Non-parametric spectral methods
• Suggests problem decomposition

12
Time Series Prediction of Load
Nonlinear
Linear
TAR
Markov
Non-parametric
Parametric
Stationary
Non-stationary
ARMA, AR, MA
Self-similar
Non-self-similar
ARIMA
ARFIMA, FGN
Best Mean
13
t1 Predictions
14
t5 Prediction
15
Conclusions

• Load has structure to exploit for prediction
• Structure is complex (self-similarity, epochs)
• Simple time series models are promising
• Benefits of more sophisticated models are
  unclear
• Current research questions
• What are the benefits of more sophisticated
  models?
• How to characterize prediction error to user?
• Is there a measure of inherent predictability?
• How to incorporate load prediction into systems?

16
(No Transcript)