based System Management

naivenorthAI and Robotics

Nov 8, 2013 (4 years and 1 day ago)

73 views

© 2002 IBM Corporation

IBM Research

1

Policy Transformation Techniques in Policy
-
based System Management

Mandis Beigi, Seraphin Calo and Dinesh Verma

mandis, scalo, dverma @ us.ibm.com

IBM Research

June 7, 2004

IBM Research

© 2003 IBM Corporation

2

Policy Transformation

Advantages


Simplifies policy
-
based management


Hides complex policies from administrators


Provides business level abstractions


Objective


Build a generic policy transformer


To be used by many disciplines


Bidirectional policy transformer


Business level to low level configuration


Low level configuration to business level (Policy Advisor)

IBM Research

© 2003 IBM Corporation

3

Policy Transformation


Types of transformation


Offline


Uses static predefined rules


Real time/online


feedback loop


Transformation taken place in 2 places


At management tool


before placed onto repository


At decision point


before send to the enforcement points




IBM Research

© 2003 IBM Corporation

4

Policy
-
based System Management



Policies

Policy

Repository

Policy

Decision Point

Policy

Enforcement
Point


Policy

Management

Tool

Policies

Actions

Policies

IBM Research

© 2003 IBM Corporation

5

Policy Transformation Enablement



Configuration

Policies

Policy

Repository

Policy

Decision Point

Policy

Enforcement
Point


Policy

Management

Tool

Policy

Transformation

Module

Goal

Policies

Actions

Policies

Case

Database

Static

Transformation

Rules

IBM Research

© 2003 IBM Corporation

6

Existing Approaches


Analytical Models

Need model of the system

Need to solve for model parameters

Need to make simplifying assumptions

Drawback
-

Exact models do not exist for real
-
life environments


Online Adaptive Control

Using concepts from control theory

Develop a neural network model

Drawback
-

Discipline specific


Simulation Approach

Model the system using a simulator

Drawback


Discipline specific

IBM Research

© 2003 IBM Corporation

7

Proposed Approach: Transformation using static rules


Example:

High Level policy:

If the
user

is from
Schwab
, then provide
Gold

level service


Low Level Policy:

If the
user

is from the subnet
9.10.3.0/24
, then reserve a
bandwidth

of
20 Mbps

and provide an
encryption of
128 bits


Transformation Rules:

1.
Schwab user

is on the

9.10.3.0/24

subnet

2.
Gold

service is to provide a
bandwidth of
20 Mbps

and an
encryption of
128 bits


IBM Research

© 2003 IBM Corporation

8

Proposed Approach: Transformation based on table
lookup

Transformation module holds a table of policies appropriate for the system



A

B

D

C

Config 1

Config 2


Hypercube for

Incoming policy


Hypercube

Representation of

Input policy

Hypercube

Representation of

Output policies

IBM Research

© 2003 IBM Corporation

9

Proposed Approach: Transformation using case
based reasoning

Applications using CBR:

1.
Diagnostics

2.
Planning

3.
Prediction


A table of cases is kept from past measurements or training set


Data my consist:


Noise


Inconsistent cases


Multiple cases having the same outcome


Missing measurements

IBM Research

© 2003 IBM Corporation

10

Sample of a Case Database

Tier0 # of
Disks

Tier1 # of
Disks

Tier0 # of
Nodes

Tier1 # of
Nodes

User
Response
Time

1

4

1

2

0
.
039

sec

3

2

2

4

0
.
029

sec

2

3

2

4

0
.
082

sec

4

1

1

2

0
.
015

sec

2

1

3

3

0
.
042

sec

2

4

1

2

0
.
053

sec

1

2

2

4

0
.
032

sec

3

2

3

4

0
.
098

sec

IBM Research

© 2003 IBM Corporation

11

Feature Selection


A typical system would have many configuration parameters and
goal values

How do you know which are the right set of configs and goals to include in
the transform data?

How does one eliminate unnecessary responses


Need to select a subset of best features from the monitored data

E1,
E2
, E3,
E4
,
E5
, E6


G1, G2, G3

Select E2, E4 and E5

Or give them different weights by sorting them according to most
relevant

to least relevant


IBM Research

© 2003 IBM Corporation

12

Feature Selection


Backward Generation

Remove one feature at a time until desired accuracy is obtained


Complete Search Strategy

Consider all possible combinations of features


Accuracy Evaluation

Measured against a set of known test
-
cases


Search

Strategy

Generation

Scheme

Evaluation

Measure

Complete


Heuristic


Non
-
deterministic

Accuracy



Consistency



Classic

Forward

Backward

Random

IBM Research

© 2003 IBM Corporation

13

Data Pre
-
processing


Removes irrelevant or redundant data to make data simpler


Reduces computational overhead


Increases accuracy


3
-
step process:

1.
Dimensionality reduction


By Feature Selection


Weights based on accuracy


Principal Component Analysis (PCA)


Combines correlated axes


Cross Correlation Matrix


Direct relationship of variables


Linearly dependent variables

Correlation:

ρ
12
= ∑ x
1
x
2
/ N
σ
1

σ
2

2.
Normalizing

3.
Data Unit Consistencies

IBM Research

© 2003 IBM Corporation

14

Principal Component Analysis


Finds components that represent maximum variance


A set of correlated variables


a set of uncorrelated variables


Reduces dimension of data


Reduces noise


Example: Linear reduction of 2 dimensions to 1 dimension

IBM Research

© 2003 IBM Corporation

15

Data Clustering


K
-
nearest neighbor clustering


Fixed or variable number of clusters


Robust to noise in data


Find the cluster with the smallest distance to lookup data point



Cluster 1

Cluster 3

Cluster 2

Config 2

Config1

Goal 3

Goal 2

Goal 1

IBM Research

© 2003 IBM Corporation

16

Experiments and Results


IBM High Volume Web Site simulator


Multi
-
tiered web site


Different workload patterns

1.
Online shopping

2.
Trading

3.
Reservations

4.
Auctions


User session characteristics


Software/hardware characteristics per tier


IBM Research

© 2003 IBM Corporation

17

Web
-
site Architecture







Web

Presentation

Server

Web

Application

Server

Database

Server


Network

IBM Research

© 2003 IBM Corporation

18

Experiment


N: # of configuration parameters


M: # of goal parameters


k: # of clusters


N=21

M=16

k=data size/100


Each case:

Generated 21 uniformly distributed random variables

Measured the goal values


Generated 100,000 data points (cases)


IBM Research

© 2003 IBM Corporation

19

Components of the Transformation Module



Cases

HVWS

Simulator

Transformation

Module

Case

Database

Data

Preprocessing

Component

Feature

Reduction

Component

Clustering

Component

Input policy

(e.g. Configuration
values)

Output policy

(e.g. Goal values)

IBM Research

© 2003 IBM Corporation

20

Accuracy of the System

Accuracy versus Data Size
N=21, M=16, k=dataSize/100
0
100
200
300
400
500
500.00
3,000.00
4,500.00
6,000.00
7,500.00
9000
20,000.00
50,000.00
90,000.00
DataSize
Distance

Euclidean distance


PCA:

M=16
-
> M’=6

reduced M by 10


In real systems N can also be reduced

IBM Research

© 2003 IBM Corporation

21

Experiment Results


Cross correlation matrix


Values between
-
1 and +1


10,000 cases


Configuration Knob

Goal value

Cross Correlation Value

‘ThinkTime’


‘SessionTime’


+0.949

‘BackgroundUtilization’

‘CPUUtilization’


+0.805

Tier0 # of
Nodes

Tier1 # of
Nodes

Tier2 # of
Nodes

Average

Response

Time

0.08

0.18

0.13

IBM Research

© 2003 IBM Corporation

22

Online/Real Time Policy Transformation



Configuration

Policies

Policy

Repository

Policy

Decision Point

Policy

Enforcement
Point


Policy

Management

Tool

Policy

Transformation

Module

Monitoring

Module

Goal

Policies

Syst
em

Data

Actions

Policies

System

Data

System

Data

Case

Database

IBM Research

© 2003 IBM Corporation

23

Online Transformation


Online monitoring component


Ensures objectives are being met


Configuration parameters are dynamically modified


Configuration parameters and objectives are measured


Builds case database


Useful for state dependent systems


IBM Research

© 2003 IBM Corporation

24

Summary


Different types of transformation


Discipline independent
-

generic solution


Offline and online methods