IBM Software Group

Predictive Analytics And

IT Service Management
Session 11479
Wednesday, August 8th
1:30 2:30 PM

Ed Woods
Consulting IT Specialist
IBM Corporation
2012 IBM Corporation

IBM Software Group | Tivoli software

Agenda

What is Predictive Analytics?

Examples

How is predictive analytics relevant to IT Service
Management?

Typical monitoring and management paradigms

Real time information versus historical data
collection

Univariate versus multivariate analysis

Examples of relevant metrics

Where to begin
2012 IBM Corporation

IBM Software Group | Tivoli software

What Is Predictive Analysis?

An area of analysis that deals with extracting information from
data and using it to predict future trends and behavior patterns

Relies on capturing relationships between explanatory
variables and the predicted variables from past occurrences

Exploit the information to predict future outcomes

Accuracy and usability of results will depend greatly on the

quality of data analysis and the quality of assumptions

Predictive analysis is used in many facets of business

Common example would be credit score

Function of many data items

Income, payment history, amount of outstanding debt, etc

2012 IBM Corporation

IBM Software Group | Tivoli software

Steps In The Predictive Analytics Process

Data organization and cleansing

Data
sources

Identify data sources

Data Mining

Analysis of data to identify underlying trends,
patterns, or relationships

Identify data to be used to develop the
predictive model

Model Development - Regression models

Data
cleansing and
organizing

Regression modeling describes the

relationship between dependent variable (the
variable to be predicted) and independent
explanatory variables

Regression models imply some level of
causation (versus correlation)

Data
mining
Model development

Predictive Model
2012 IBM Corporation

IBM Software Group | Tivoli software

Predictive Analytics
About Regression Models And Types Of Models

Regression models are the core of predictive analytics

A wide variety of models can be applied

Linear regression model

Analyzes the relationship between the response or dependent variable
and a set of independent or predictor variables

Partial or Stepwise regression

Modeler does not specify all the explanatory variables

Variables are added iteratively

Logit or Probit regressions

Allow one to predict a discrete outcome (yes/no) from a set of variables

Time series models

Used for predicting or forecasting the future behavior of variables

Data points taken over time may have an inherent time relation

Developed to decompose the trend, seasonal and cyclical component of
the data

Many more models
2012 IBM Corporation

IBM Software Group | Tivoli software

Examples Of Predictive Analytics Commonly

Applied to IT

Performance modeling

z/OS workload right sizing and load balancing

Model workload placement using SMF data as input

Trending and forecasting of workload/resource utilization

Workload performance trends

Discern patterns in resource utilization

Capacity planning

The common question >> When will a critical resource reach breaking point?

What If Analysis examples

DB2 buffer pool analysis

DB2 performance trace data to determine optimal pool sizing and object placement

DB2 SQL and object tuning

DB2 Explain analysis based on DB2 Catalog statistics and SQL call changes

2012 IBM Corporation

IBM Software Group | Tivoli software

A Goal For Many Shops

Make Systems Management More Proactive

In many shops systems management tends to be done ad hoc

Some alert generation varies by shop

Some shops very alert driven many are not

Often notification consists of call the help desk

Many customers want to be more proactive

Definition of proactive may vary

Proactive for some installations may mean more rapid alert and
notification of technical and/or business application issues

Proactive for some installations may mean notification prior to the
problem
Alert when utilization indicates a potential issue in the future
Alert when Im within 90% of the wall

2012 IBM Corporation

IBM Software Group | Tivoli software

The Typical Monitoring Paradigm

Traditional monitoring strategy

Monitor key resources based upon established best practices

Resource utilization and resource bottlenecks

Monitor performance and availability

Key Performance Indicators (KPIs)
Examples Response time, transaction rate, technical component,
software subsystem, or business application availability

Monitor based on established SLAs

Alert notification about performance bottlenecks and outages

Notification via monitoring UIs, paging, emails

Real time monitoring versus historical

Real time monitoring for current utilization and status

Historical data collection for trending and after the fact analysis

Most shops monitor but how predictive is it?

2012 IBM Corporation

IBM Software Group | Tivoli software

Real Time Monitoring Provides A Starting Point

For Analysis
Real time monitoring provides a view
of current utilization, status, and alerts

Data
CICS Response time

Alerts

Provides a view of current status, but is

not necessarily predictive in nature

DB2 Distributed threads

DB2 network

CICS network

Alerts
IMS Response time

IMS network

Commands
2012 IBM Corporation

IBM Software Group | Tivoli software

Historical Data Analysis

Helps Identify Critical Metrics, Trends, Usage Patterns And Potential Issues

Real time

History snapshot data

History plot chart

Average CPU over the past 12 hours

2012 IBM Corporation

10

IBM Software Group | Tivoli software

Another Example - Historical Baseline Data To

Compare Past Trends To Current Trends
Request yesterdays data

Uses detailed data

Yesterday

Today

Example compare
todays CPU utilization
trend to yesterday

2012 IBM Corporation

11

IBM Software Group | Tivoli software

The Problem:
Traditional Monitoring Approaches Have Limitations

Many tools, data sources and metrics available

Many are Resource/Single Metric Focused (Univariate)

Often many missed, or misinterpreted events

In many shops not enough time, and/or resources to correlate
completely

May require many people and groups to collaborate effectively

Many resources and no obvious resource inter-relationships
Univariate - refers to an expression, equation,
function or polynomial of only one variable

Multivariate - encompasses the

simultaneous observation and analysis
of more than one statistical variable
2012 IBM Corporation

12

IBM Software Group | Tivoli software

Problem Analysis And Resolution

In many IT environments

Problem identification and notification may be ad hoc

Alert notification via phone calls, emails, or paging

Problem analysis is often after the fact

Problem analysis and resolution often involves rounding

up the usual suspects (and getting them to confess)

Issue resolution relies heavily on the knowledge and
intuition of the technical staff

Knowledge of the systems and business applications

Understanding complex problems will be multivariate in nature

2012 IBM Corporation

13

IBM Software Group | Tivoli software

Why Multivariate Analysis?

Multivariate analysis expands the relevance of the predictive
analytic approach

Provides context through correlation

Example credit rating metrics

Payment history how relevant if I do not consider other metrics?

Income again how relevant if I do not consider other metrics?

Multivariate is important for IT Service Management

Many business applications are composite in nature

Many components, platforms, core technologies

Many critical resources are shared and inter-related

Mainframes support many applications

Networks may support a wide array of workloads

2012 IBM Corporation

14

IBM Software Group | Tivoli software

An Example Of Multivariate Analytics

Goal -> Predict/Identify Issues (Early
Warning)

Analyze performance data and combine
Univariate with Multivariate Analytics

Predict: Eruption
forecasting using
seismic energy..

Identify Metric Inter-relationships

Detect anomalies rapidly, as metrics
deviate from normal behaviour individually
and from the correlated group

2012 IBM Corporation

15

IBM Software Group | Tivoli software

Using Predictive Analytics To Expand The Typical

Analysis Paradigm

Univariate metrics may be useful for some forms of analysis

Many typical IT challenges are multivariate in nature

Most applications are multi-component and multi-layered

Most applications cross software platforms and may involve multiple hops

Complete analysis may require metrics from multiple sources

Exploit the knowledge and expertise of the IT staff to begin to form a

multivariate approach

Many may use predictive analytics in an informal manner

Proven knowledge of business applications (school of hard knocks)

Knowledge as documented in established industry best practices

2012 IBM Corporation

16

IBM Software Group | Tivoli software

How Does Multivariate Analytics Detect Problems Sooner?

Server Memory (%)
Alert
(90%)
Normal
Range

Alert!

Multivariate analytics detects problems sooner by

detecting the deviation of metrics that normally move
together.
For example:
Memory consumption is normally correlated to
HTTP requests

Memory
(%)

Static Threshold = Short Warning

Server Memory + HTTP Requests

But when memory deviates from HTTP Requests, as

would happen with a memory leak, this indicates a
problem and an alert is generated.
The alert is generated much sooner than waiting for a
static threshold violation.

This advanced warning time helps you become

proactive and mitigate damage before customer
Alert! service is impacted.

Normal
Range
Memory
(%)

It also help reduce threshold alerts due to normal

threshold violation correlated with HTTP Requests.

HTTP
Requests
Multivariate = Alerts earlier on Deviation

2012 IBM Corporation

17

IBM Software Group | Tivoli software

Examples Of IT- related Multivariate Metrics

DB2 example

DB2 object lock conflict >>

long running SQL call >> high In-DB2 time >> longer thread
elapsed time >> longer DB2 query time

IMS example

High IMS message region occupancy time >>

IMS transactions queued >> longer IMS transaction scheduling
time >> longer IMS response time >> lower IMS transaction
processing rate

MQ example

Lower MQ message input rate >>

Higher MQ message queue depth >> lower transaction
processing rate >> longer CICS/IMS transaction response time
2012 IBM Corporation

18

IBM Software Group | Tivoli software

An Example Of Multivariate Analysis For IMS Performance

Monitor and trend multiple IMS
performance metrics over time

Problem transaction
count by status

Response time and processing rate

Plot chart analysis of key

IMS performance metrics

IMS Bottlenecks

Enqueue/dequeue rates
2012 IBM Corporation

19

IBM Software Group | Tivoli software

Identifying The Critical Metrics

Knowledge of business applications

Internal operational processes

Known issues based upon prior operational experience

Maintaining a history of common alerts/events

Identify critical performance metrics as established by best

practices documented in commonly available sources

IBM documentation and IBM Red Books

Share, CMG, IDUG, Pulse, IOD and other user group presentations

Define a list of the most critical metrics to track

Consider each component/platform for the application(s)

Consider various data sources

Monitoring, automation, console logs, application data sources

2012 IBM Corporation

20

IBM Software Group | Tivoli software

Example Critical IMS Performance Metrics

CICS
Threads

Connection
bottlenecks

Network
delays

IMSConnect
Connect
IMS
APPC
APPC
Message In
OTMA
OTMA
Telnet
Telnet

App Init &

execution

IMS
Control
Region
Queues &
Scheduling

IMS
Message
& BMP
Regions

Network
delays

IMS
Control
Region

Message Out

Threads
DB, BP
I/O delays

Connection
bottlenecks

Threads
DB2 Subsystem
DB, BP
I/O delays

IMS
DLI

Lock
Conflicts

IRLM

Lock
Conflicts
2012 IBM Corporation

21

IBM Software Group | Tivoli software

Another Example Typical DB2 Performance Metrics

DB2 Performance Metrics
DB2 thread connection issues
Long running threads
DB2 SQL delays and long running SQL
DB2 database I/O delays and BP perf.
Network and/or OS delays
DB2 lock conflicts
Connect GW
bottlenecks

DB2 App

DB2 App

DB2 Connect

Network
delays

Network
delays

z/OS
IMS

MSTR

CICS
Batch

Connection
bottlenecks

Connections
Threads
Logging

TSO
WebSphere

IRLM
LockLocking
Conflicts

Connection
bottlenecks

DBM1

DDF
SQL

SQL
EDM
PTs, SKPTs, CTs, SKCTs
DSC Dynamic SQL
Sort Pool

RID Pool

DB, BP
Buffer
Mgr - VPs
I/O delays

Distributed
threads
SP Addr Space(s)
Stored Procedures
UDFs

SP Sched
Delays

2012 IBM Corporation

22

IBM Software Group | Tivoli software

Other Examples Of Common z/OS Critical

Performance Metrics
WebSphere MQ

CICS

Queue depth
Message send/receive rate
DLQ depth
Channel status and performance

Transaction response time

Transaction rate
Region CPU rate
File I/O count
String waits
Abend messages

z/OS
System CPU rate
Paging rate
WLM Performance Index
DASD I/O MSR time and rate
Critical console messages

WebSphere
Method call count and elapsed time
Heap size
Garbage collection
Connection pool utilization

Network
Network Connection status and performance
Network interface utilization
2012 IBM Corporation

23

IBM Software Group | Tivoli software

Dont Overlook Alerts

Alerts Can Provide Valuable Metrics
Alerts may be a useful source of metrics for analysis
Number of alerts and frequency of alerts may be useful
Correlate alerts to identify critical metrics

2012 IBM Corporation

24

IBM Software Group | Tivoli software

Predictive Analytics
Defining A Strategy Whats Required

A predictive analytics approach starts with a comprehensive historical
data collection strategy

Gather history to a common point, if technically feasible

Consider historical collection challenges carefully

Different source platforms (example - LUW versus z/OS or z/VM)

Different instrumentation, trace options, data formats

z/OS SMF, DB2 logs, DB2 trace IFCIDs, CICS SMF records, IMS logs,
WebSphere PMI trace interface, Windows performance counters, and
much more.

Collection frequency and quantity

Avoid flooding with useless information

Focus on quality versus quantity

What metrics to collect

What quantity of data

Where to house the data

What level of detail and/or summarization
2012 IBM Corporation

25

IBM Software Group | Tivoli software

Predictive Analytics Begins With History

Historical Data Collection Considerations
Cost
Of Collection

Diagnostic
Value

Monitoring And History Collection Trade-off

Historical data collection varies in cost and quantity

CPU, memory, and software process cost of collection

Cost of data storage and retention

Cost of retrieval and post processing

Ease of review and analysis

Some historical data will be more relevant and useful than other data

Consider the context, nature, and meaningfulness of the data
2012 IBM Corporation

26

IBM Software Group | Tivoli software

Types Of Historical Monitoring Data

Know the nature and characteristics of the history data being collected

Detail data

Data that documents/measures detail of a specific event

Often high quantity data and the most detailed for analysis

May pose the greatest challenge in terms of cost, retention, post processing

Examples DB2 Accounting records, CICS SMF 110 records, IMS log records

Summary data

Data that summarizes underlying detail data

Either an aggregation or an averaging of underlying detail records

May be useful for longer term trending and analysis

Reduces quantity of data and reduces cost of retention, post processing

Less detail may mean less diagnostic value

2012 IBM Corporation

27

IBM Software Group | Tivoli software

Types Of Historical Monitoring Data - continued

Interval data

History data that includes an encapsulation of one or multiple events to a specified
time interval

The data will include all activity within that given time interval

Useful for problem analysis and trending analysis

Examples DB2 statistics records

Snapshot data

Typically a point in time snapshot of activity

Snapshots are usually based on a specified time interval

Snapshots may be taken of types of history (detail, summary, or interval)

Snapshots will show activity at time of the snapshot, but may/may not reflect
activity between snapshots

Useful for problem analysis and trending analysis

Useful as an aid in setting alert thresholds

Examples snapshot history captured by performance monitoring,
2012 IBM Corporation

28

IBM Software Group | Tivoli software

Predictive Analytics
Defining A Strategy Whats Required - continued

Pursue a multivariate approach where feasible

Univariate is useful for certain trending and modeling uses

Multiple variables provide a more meaningful indication of potential
issues

Multivariate is more relevant to todays composite applications

Multivariate poses challenges

How to best identify the key metrics to collect

How much data to collect

How to analyze and correlate the information

How to display the data correlation result

How to feed the result to other systems

Monitoring, automation, business application views, help desk views

2012 IBM Corporation

29

IBM Software Group | Tivoli software

Predictive Analytics Summary

Whats Needed

Identify information metric sources

Monitoring tools, platforms, alerts, console and application logs

Define metrics

What are the most critical metrics to track?

History

Define a collection strategy that allows for the aggregation of data

Correlation methodology

How to correlate metrics in real time and in history

Is there a way to automate the correlation process?

Display and analysis methodology

How to analyze and display critical metrics data and alerts

Prediction

How predictive are the chosen metrics?
2012 IBM Corporation

30

IBM Software Group | Tivoli software

Thank You!!

2012 IBM Corporation

31

IBM Software Group | Tivoli software

Check Out My Blog

http://tivoliwithaz.blogspot.com

2012 IBM Corporation

32