An Adaptive Middleware for Supporting TimeCritical Event Response

1
An Adaptive Middleware for Supporting
Time-Critical Event Response

• Qian Zhu Gagan Agrawal

2
Motivating Application Great Lake Nowcasting and
Forecasting
3
Motivating Application Real-Time Volume Rendering
4
Problem Description

• Task
• Intensive computation and communication
• Application-specific flexibility
• Constraints
• Tightly bounded time constraints
• Resource availability detected at runtime
• Goal
• To optimize a benefit function

5
Middleware Design Goals

• To enable the time-critical event handling to
  achieve the maximum benefit, while satisfying the
  time constraint
• To be compatible with Grid and Web Services
• To enable easy deployment and management with
  minimum human intervention
• To be used in a heterogeneous distributed
  environment

6
Overall Design
7
Adaptation Algorithm

• In the normal processing phase
• For each parameter X in service S
• For each checkpoint
• Learn the relationship of the value of X and the
  execution time
• Learn the relationship of the value of X and gain
  from the benefit function
• Update the system model
• In the time critical event handling phase
• Adjust X based on the system model
• Accelerate the adaptation if violating the time
  deadline

8
Experimental Evaluation

• Demonstrate that the service parameters converge
  quickly while meeting the time constraint.
• Demonstrate that the overhead of adaptation is
  modest.
• Demonstrate the overhead caused by learning is
  very small.

9
Application

• Great Lake Forecasting System
• Dataset
• Grid resolutions 2000m2000m, 1000m1000m and
  500m500m
• Service Parameters
• Grid resolutions
• External time step, internal time step
• Benefit Function

10
Task

• Time constraint 2 hours
• To forecast the meteorological condition of a
  certain region in the Lake Erie for the next two
  days
• Result
• Coarse grid resolution 2000m2000m
• Fine grid resolution 500m500m
• Outputs water level, water temperature, UM, VM,
  and W

11
External Time Step
12
Internal Time Step
13
Grid Resolution
14
Overhead of the Adaptation Algorithm
time-critical event handing phase
14
12
10
15
Overhead of the Adaptation Algorithm learning
phase

• The overhead of the adaptation algorithm for
  tuning 1,2 and 3 parameters is 4.1, 7.5 and
  10.1.

16
Application

• Volume Rendering
• Dataset
• 7.5GB(block size 32321664KB)
• 30 time steps
• Service Parameters
• Error tolerance, image size
• Wavelet coefficients
• Benefit Function

17
Task

• Time constraint 20 minutes
• Initial view angles 45, 90, 135
• Image size 256256,
• error tolerance lt 0.03
• Result
• Actual view angles 45, 75, 90, 135
• Error tolerance 0.02
• Image size 256256

18
Error Tolerance
19
Image Size
20
Wavelet Coefficient
21
Overhead of the Adaptation Algorithm
time-critical event handing phase
12
11
9
22
Overhead of the Adaptation Algorithm learning
phase

• The overhead of the adaptation algorithm for
  tuning 1,2 and 3 parameters is 2.2, 3.0 and
  4.8.

23
System Model

• State Equation

24
System Model Contd
adaptation overhead

• Performance Measure

benefit
time constraint

• Constraints

Note Each service is deployed on a single node
25
Initial Policy

• It is simple and straightforward
• Parameter convergence depends on the learning
  rate
• It may occur a large adaptation overhead