Presentation on the work executed by Stefano Doni of Moviri regarding the analysis of Oracle Workloads using deeper characterization than the common "by instance" or "by process" technique

1. CCMG 2014
Cromwell (CT) – April 25th, 2014
Renato Bonomini – renato.bonomini@moviri.com
Andrea Vasco – andrea.vasco@moviri.com
Stefano Doni – stefano.doni@moviri.com
Demystifying Oracle database
capacity management with
workload characterization

2. 2 CCMG - Connecticut Computer Measurement Group 2014
§ Business-aware Capacity Management in a nutshell
§ Database modeling: common pitfalls
§ The workload characterization methodology
§ How to do it for Oracle databases
§ Results and Conclusions
Agenda

3. 3 CCMG - Connecticut Computer Measurement Group 2014
The Capacity Modeling a generic application can be approached focusing on different yet incremental
scopes; ITIL defines a set of sub-processes such as:
§ Component Capacity Management: manage, control and predict the performance, utilization and
capacity of IT resources and individual IT components
§ Business Capacity Management: translate business needs and plans into capacity and performance
requirements for services and IT infrastructure, to ensure that future capacity and performance needs
can be fulfilled
Capacity Management Approaches
HTTP
requests
Web
resources
Application
Components
(servlet)
Business
Logic
(java beans)
Database
instances
Web
transaction
Silo reporting
BA Approach

4. 4 CCMG - Connecticut Computer Measurement Group 2014
Component Capacity Management cannot:
§ Define and explain the capacity of a service in
business terms (e.g. number of users)
§ Identify which component is the current
bottleneck in the application
§ Optimize capacity at the application level (e.g.
moving capacity between different layers)
Component capacity management sometimes is not enough
OS & Hardware
Middleware
Applications
Services
?
Will batch processes end timely?
Is there capacity to manage +10% users?

5. 5 CCMG - Connecticut Computer Measurement Group 2014
Why do we need Business-Aware Capacity Management?
The goal is to express capacity in “business terms” which are
1. easy to understand and to share with stakeholder (Business Service Owners, Marketing, etc.)
2. Independent from the underlying technology and its changes
Q: Is this correlation always easy to find?

6. 6 CCMG - Connecticut Computer Measurement Group 2014
§ Front-end layers can be simply modelled using single business KPIs
− they are specialized for some well-defined business functions (e.g. application servers)
− the resource utilization is usually not workload dependant (e.g. HTTP request)
− the workload composition is stationary
Front-end layers? Yes, it’s generally easy
R² = 0.97831
0%
20%
40%
60%
80%
100%
0 500 1000 1500 2000 2500 3000 3500 4000
PhysicalCPUconsumed
Purchase Orders Volume / Hour
Application Server
Physical CPU consumed vs Purchase Orders Volume

7. 7 CCMG - Connecticut Computer Measurement Group 2014
§ Databases are inherently shared and their resources pulled in several directions
− By multiple functions within the same application: the same database is typically used for both
online transactions, background batches, business reporting, etc.
− By multiple applications accessing the same database: quite often a single database is accessed by
multiple applications at the same time, for different purposes
− By infrastructure consolidation: a single server might run multiple database instances
Database layers? The previous “simple” modeling would not fit…
R² = 0.51894
0
2
4
6
8
10
12
0 500 1000 1500 2000 2500 3000 3500
PhysicalCPUconsumed
Purchase Orders Volume / hour
Database Server
Physical CPU consumed vs Purchase Orders Volume

8. 8 CCMG - Connecticut Computer Measurement Group 2014
Agenda
§ Business-aware Capacity Management in a nutshell
§ Database modeling: common pitfalls
§ The workload characterization methodology
§ How to do it for Oracle databases
§ Results and Conclusions

9. 9 CCMG - Connecticut Computer Measurement Group 2014
How the workload correlation usually degrades while traveling the different layers
0
0.5
CPUcores#
Application Server
Physical CPU consumed
0
10
20
CPUcores#
Database Server
Physical CPU consumed
0
1000
2000
3000
4000
#Request
Frontend Server
Number of Request
HTTP requests
Front-end
Database layer
…

10. 10 CCMG - Connecticut Computer Measurement Group 2014
§ Single workload models for databases typically do not explain resource utilization
§ Statistically poor models can not be used for prediction purposes
What does it mean? “Simple” models do not work for databases
R² = 0.51894
0
2
4
6
8
10
12
0 500 1000 1500 2000 2500 3000 3500
PhysicalCPUconsumed
Purchase Orders Volume / hour
Database Server
Physical CPU consumed vs Purchase Orders Volume
R² = 0.97831
0%
20%
40%
60%
80%
100%
0 500 1000 1500 2000 2500 3000 3500 4000
PhysicalCPUconsumed
Purchase Orders Volume / Hour
Application Server
Physical CPU consumed vs Purchase Orders Volume

11. 11 CCMG - Connecticut Computer Measurement Group 2014
Agenda
§ Business-aware Capacity Management in a nutshell
§ Database modeling: common pitfalls
§ The workload characterization methodology
§ How to do it for Oracle databases
§ Results and Conclusions

12. 12 CCMG - Connecticut Computer Measurement Group 2014
§ The Utilization Law (Little’s Law) describes the relationship between workload intensity and resource
utilization
Leveraging the Utilization Law
U = Xn
Sn
n
∑ +U0
U resource utilization
U0
residual utilization
Xn
processed work (throughput)
Sn
service demand
§ The Utilization Law can be extended to include any number of workload classes; in the database
modeling context, a workload class can represent a specific source of work we need to model:
U =X1
S1
+X2
S2
+....+Xn
Sn
+U0
Q: how can we derive the service demand of each workload class?

13. 13 CCMG - Connecticut Computer Measurement Group 2014
Estimate Si using U and Xi
§ Multivariate regression models to
estimate service demands
§ Required when only the total resource
utilization is known
§ Drawbacks
− Not always feasible due to statistical
constraints (e.g. multicollinearity)
− Lack of visual interpretation
− Difficult to understand
Direct Measurement of Workload Demand
§ The single workload resource
consumption can be measured
− E.g. we know how much CPU is used for
business reporting, online transactions,
receipt batches etc.
§ It is possible to manage each workload
independently
Two service demand modeling approaches
“Divide et Impera”
This is the approach we can use
U =X1
S1
+X2
S2
+....+Xn
Sn
+U0
then Si
= ?

14. 14 CCMG - Connecticut Computer Measurement Group 2014
Agenda
§ Business-aware Capacity Management in a nutshell
§ Database modeling: common pitfalls
§ The workload characterization methodology
§ How to do it for Oracle databases
§ Results and Conclusions

15. 15 CCMG - Connecticut Computer Measurement Group 2014
0
2
4
6
8
10
12
CPUcores#
Database Server
Physical CPU consumed
Estimating resource utilization for each workload class
The selected approach relies on the ability to identify all of the different components
contributing to the measured resource utilization
Oracle itself provides this information!

16. 16 CCMG - Connecticut Computer Measurement Group 2014
§ Oracle automatically tracks detailed resource consumption metrics, which are stored in
circular buffers held in database memory and later flushed to disk for historical auditing
§ Metrics can be accessed with standard queries estimating resource consumption:
− Per session
− Per SQL Statement
How to estimate the contribution on resource utilization for each component
It’s possible to slice resource consumptions at different levels:
it depends on how the application works!
§ The identification of the impact of a workload can be derived for instance by considering:
− the specific user the application is using
− the machine generating the request

17. 17 CCMG - Connecticut Computer Measurement Group 2014
§ Metrics are stored into:
− V$ACTIVE_SESSION_HISTORY view for about 1 hr
− DBA_HIST_ACTIVE_SESS_HISTORY table for 7 days
§ Statistical analysis of sessions:
− Session State (ON CPU, WAITING)
− Wait Class, Event (User I/O, file scattered read, …)
− Application (Program, Module, Action, …)
− Number of IO operations / bytes transferred
during the interval
§ How to aggregate the application sessions to get
the total number of CPU time consumed?
− count the concurrent number of sessions
(the ones that are ‘ON CPU’)
− sum all the CPU times within the desired time
window (i.e. one hour), getting the total CPU busy
time in seconds
Resource consumption components separation: an estimation method
SELECT TO_CHAR(sample_time, 'yyyy/mm/dd
HH24') as sample_time,
username,
round(sum(DECODE(session_state,'ON CPU',
DECODE(session_type,'BACKGROUND',0,1),
0))/60/60*10,2) AS cpu_secs
FROM DBA_HIST_ACTIVE_SESS_HISTORY a,
dba_users b
WHERE sample_time > sysdate – 7
and a.user_id=b.user_id
GROUP BY
TO_CHAR(sample_time, 'yyyy/mm/dd HH24'),
username

18. 18 CCMG - Connecticut Computer Measurement Group 2014
Agenda
§ Business-aware Capacity Management in a nutshell
§ Database modeling: common pitfalls
§ The workload characterization methodology
§ How to do it for Oracle databases
§ Results and Conclusions

19. 19 CCMG - Connecticut Computer Measurement Group 2014
Workload characterization is an essential step in capacity management: if properly conducted, it
can provide remarkably good models for complex contexts.
Single workload models work for Oracle Databases
R² = 0.96666
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1.8
2.0
0 500 1000 1500 2000 2500 3000 3500
EstimatedPhysicalCPUconsumed
Purchase Order Volume / hour
Database Server
Estimated Physical CPU consumed vs. Purchase Orders Volume

20. 20 CCMG - Connecticut Computer Measurement Group 2014
Questions?

21. 21 CCMG - Connecticut Computer Measurement Group 2014
Thank you!

22. US Office
One Boston Place, Suite 2600
Boston, MA 02108
T: +1-617-936-0212
Headquarters
Via Schiaffino, 11
20158, Milan Italy
T: +39 02 4951 7001
@moviri
moviri
moviri