Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

download Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

of 48

Transcript of Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    1/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 1

    Capacity Management for Oracle

    Exadata Database Machine v2

    Dr. Boris Zibitsker, BEZ Systems

    [email protected]

    www.bez.com

    OOW 2010

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    2/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 2

    About the Author

    Dr. Boris Zibitsker, Chairman, CTO, BEZ Systems .

    Boris and his colleagues developed modeling technology supporting

    multi-tier distributed systems based on Oracle, Teradata, DB2, and

    SQL Server. Boris consults, and speaks frequently on this topic at

    many conferences across the globe.

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    3/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 3

    Focus

    Oracle Exadata Database Machine supports mix OLTP and data

    warehouse workloads, which provides a lot of benefits but require

    effective capacity management

    Capacity management includes workload management, performance

    management and capacity planning Knowing you workloads profiles and setting realistic SLOs for each

    workload is critical for Exadata DBM capacity management

    Everything is interdependent and workload growth and any change

    can improve performance of one of the workloads, but negatively

    affect response time of others

    We will review a case study illustrating capacity management solutions

    for Oracle Exadata Database Machine supporting mix workload

    growth, new applications implementation and server consolidation

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    4/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 4

    Oracle Exadata Database Machine v2

    Supports Mixed Workloads

    Each workload use one or

    several VMs, JVMs and

    Application servers

    Each workload use one or

    several RAC nodes Data spread across all

    Exadata Cell Disks

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    5/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 5

    Capacity Management Functions

    CapacityManagement

    StrategicCapacityPlanning

    TacticalPerformanceManagement

    OperationalWorkload

    Management

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    6/48

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    7/48Boris Zibitsker, BEZ Systems Predictive Analytics for IT 7

    RAC CPU Service

    RAC Delay

    RAC CPU Wait

    Oracle Exadata DBM Response Time

    Components

    Exadata CPU Wait

    InfiniBand

    Exadata CPU Service

    Exadata Disk Wait

    Exadata Flash Cash RT

    Exadata Disk Service

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    8/48

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    9/48

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    10/48 Boris Zibitsker, BEZ Systems Predictive Analytics for IT 10

    Decision Support Techniques

    Gut feelings Rules of thumb

    Regression analysis

    Analytical models

    Simulation models

    Benchmarks

    Workload Growth

    ResponseTime SLM

    SLOUti l izat ion Law:

    U = A * S , where U Utilization,A Arrival Rate,

    S Service TimeResponse Time Law:

    R = S / (1 U) , where R Response Time. See [1,2]

    SLM

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    11/48 Boris Zibitsker, BEZ Systems Predictive Analytics for IT 11

    Closed Queueing Network Model of Exadata DBM

    with Mix Workloads(many details are not shown)

    1

    2

    n

    CPU

    Memory

    1

    2

    n

    CPU

    Disk

    Flash

    Active Active

    Exadata Storage Server Grid

    Users

    Requests

    75

    60

    1550

    25

    25

    RAC DB Server Grid

    Disk

    CPUCPU Max?Max? CPU

    InfinibandSwitch

    Network

    Workloads

    This is a simplified high level analytical Queueing Network Model of the Exadata DBM

    Many details are omitted for illustration purpose

    It is a multi-tier model, where 1st tier represents RAC nodes and 2nd tier represents Exadata cells

    The number of tiers in the model can be expanded to represent middle tier application servers

    Exadata Cells, physical and flash disks, channels, infiniband, etc are represented in the model by

    the network of servers and queues

    Measurement data characterizing utilization of RAC nodes and Exadata cells can be extracted from

    Oracle OEM

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    12/48Boris Zibitsker, BEZ Systems Predictive Analytics for IT 12

    Case Study

    1. What will be the impact of the workload and volume

    of data growth?2. What will be the impact of new application

    implementation?

    3. How to justify server consolidation.

    4. What should be tuned?

    5. What is the optimum level of workload concurrency?6. What is the optimum workload priority?

    7. What is the minimum hardware upgrade required to

    support SLOs?

    8. What will be the impact of changing number of

    processors per RAC node?9. How to coordinate configuration planning for middle

    tier and Exadata Machine to support SLOs for major

    workloads.

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    13/48Boris Zibitsker, BEZ Systems Predictive Analytics for IT 13

    1. What will be the impact of the workload

    and volume of data growth?

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    14/48Boris Zibitsker, BEZ Systems Predictive Analytics for IT 14

    Oracle Exadata Database Machine v2

    Environment

    Know profile of each workload

    Analyze cyclical pattern of

    resource utilization by each

    workload

    Find representative, peakmeasurement interval as a base

    for further analysis

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    15/48Boris Zibitsker, BEZ Systems Predictive Analytics for IT 15

    RAC Nodes Performance Analysis

    Analyze seasonal trends

    Find representative and peak

    RAC CPU utilization by each of

    the workloads

    Build profile for each workload

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    16/48Boris Zibitsker, BEZ Systems Predictive Analytics for IT 16

    Exadata Cells Performance Analysis

    Exadata Cell Flash Disk RT is 1 1.5 ms

    Exadata Cell Physical Disk RT is 5 10 ms

    Exadata CPU Utilization is 5 20%

    0.000

    5.000

    10.000

    15.000

    1 2 3 4 5 6

    Average Cell Disk RT (ms)

    Average RT (ms)

    0.000

    0.500

    1.000

    1.500

    2.000

    1 2 3 4 5 6 7 8

    Average Flash Disk RT (ms)

    Average RT (ms)

    0.0

    10.0

    20.0

    30.0

    Cell CPU Util %

    Cell CPU Util %

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    17/48Boris Zibitsker, BEZ Systems Predictive Analytics for IT 17

    Predicting Workload and Volume of Data

    Growth Impact on Response Time

    Prediction shows how workload

    and volume of data growth will

    increase contention for systems

    resources and how it will affect

    RT of each of workloads

    Find when RT will not meet SLO

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    18/48Boris Zibitsker, BEZ Systems Predictive Analytics for IT 18

    Waiting Time for CPU Will Become the Largest

    Component of the Sales Response Time

    Find when SLO will not be

    met

    Find which workload will use

    most of CPU resources

    Identify options how to

    improve performance

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    19/48Boris Zibitsker, BEZ Systems Predictive Analytics for IT 19

    2. What will be the impact of new

    application implementation?

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    20/48Boris Zibitsker, BEZ Systems Predictive Analytics for IT 20

    New Application

    What will be the impact of implementing a new

    workload?

    Test environment can usedifferent hardware, software and

    DBMS platforms

    Workload profile in production

    environment will be different

    New workload will increasecontention for resources and

    affect current workloads

    performance

    Production DB MachineStress Testing

    New Appl

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    21/48Boris Zibitsker, BEZ Systems Predictive Analytics for IT 21

    NewAppl Implementation Will Impact

    Performance of Existing Workloads

    Simulation of moving workloads

    from test to production system

    predict how new workload will

    affect performance of the existing

    workloads

    Model take into consideration

    differences between hardware

    and software platforms,

    differences in volume of data, etc.

    Set realistic expectations and

    justify what should be changedproactively

    Predict how new

    application will affect

    performance of

    existing applications

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    22/48

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    23/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 23

    CPU Utilization Increase After

    Implementation of New Application

    New workload will use more

    CPU resources than existing

    workloads

    Identify proactive

    performance managementmeasures

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    24/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 24

    3. How to justify server consolidation

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    25/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 25

    How to justify server consolidation

    Production DB Machine

    WKL2

    WKL1

    WKL3

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    26/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 26

    How Server Consolidation Will Affect

    Existing Workloads

    Prediction results evaluate the

    impact of planned server

    consolidation on Oracle

    Database Machine Exadata v2

    Shows when system will not

    meet SLOs

    Identify the minimum upgrade

    required to support SLOs

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    27/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 27

    Predict Response Time for Workloads WKL1 and

    WKL2 and CPU Utilization After Consolidation

    Prediction results show how

    workloads WKL1 and WKL2 will

    perform after server

    consolidation and how it will

    affect CPU utilization of Oracle

    Database Machine Exadata Savings in power consumption,

    software licenses, maintenance,

    vs coexistence on one platform

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    28/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 28

    4. What should be tuned?

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    29/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 29

    Predicted Impact of Data Compression on

    Workloads Response Time and Disk Utilization

    Data is stored by column andthen compressed

    Factors affecting acompression ratio:

    Table size

    Data cardinality

    Read/write ratio

    Prediction results show thatdata compression affectsOLTP and DSS workloadsperformance differently

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    30/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 30

    Prediction results can be used to evaluate

    options and find solution satisfying SLOs of

    major workloads Each ASM Disk Group has

    different performance

    characteristics and cost

    Each table has different size,

    frequency of accesses by

    different workloads and pattern of

    using data

    Modeling can be used to

    evaluate different alternatives of

    placement data and find solutionthat will help to meet SLOs for

    major workloads

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    31/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 31

    Exadata Parallel Data Access Change Index

    Strategy

    Parallel access to data

    distributed across all disks

    reduce I/O service time

    Usage of Flash cache and

    flash disks, smart scanfiltrates data

    It affects decisions when to

    create indexes

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    32/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 32

    5. What is the optimum level of

    workload concurrency?

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    33/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 33

    Limiting Concurrency Reduces Contention but

    Increases # of Requests Waiting for the Thread

    Limiting Concurrency for the

    workload can reduce

    contention for resources

    Requests of the workload with

    limited concurrency will spendless time waiting for

    resources, but spend more

    time waiting for the thread

    Performance of the workload

    with limited concurrencymight suffer, but other

    workloads can have

    significant performance gain

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    34/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 34

    Predicting impact of lowering the level of

    concurrency for ETL workload

    ETL use a lot of resources,

    but satisfy SLO

    What if we limit ETL

    concurrency starting Period

    #3? ETL time to load data will

    increase, but will be

    satisfactory

    Response time for other

    workloads will improve

    0

    0.2

    0.4

    0.6

    0.8

    1

    1.2

    1.4

    1 2 3 4 5 6 7 8 9 10 11 12

    Relative Response Time

    Sales

    Marketing

    HR

    Archive_1

    ETL

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    35/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 35

    6. What is the optimum workload

    priority?

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    36/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 36

    What will be an impact of workload priority

    change?

    Increase Priority for the

    critical Workloads will

    Improve their performance

    but negatively affect others

    Prediction results evaluatedifferent alternatives and

    provide valuable information

    to justify proactive decisions

    0

    0.2

    0.4

    0.6

    0.8

    1

    1.2

    1.4

    1.6

    1 2 3 4 5 6 7 8 9 10 11 12

    Relative Response Time

    Sales

    Marketing

    HR

    Archive_1

    ETL

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    37/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 37

    7. What is the minimum hardware

    upgrade required to support SLOs?

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    38/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 38

    Predicted Impact of Adding 4 RAC Nodes and 14

    Exadata Cells in May 2011 on Workloads

    Response Time

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    39/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 39

    Impact of Adding 4 RAC Nodes and 14

    Exadata Cells in May 2011 on CPU and Disk

    Utilization Hardware upgrade reduce

    contention for RAC CPU

    resources and Exadata Disk

    utilization

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    40/48

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    41/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 41

    Predicted Impact of Changing Number of

    Processors per RAC Node by 50% in April 2011(assuming that Oracle will introduce new RAC node with more CPUs per node)

    Increase RAC node capacity

    will have positive impact on

    response time and reduce

    CPU utilization

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    42/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 42

    9. How to Coordinate Configuration Planning

    for Middle Tier and Oracle Exadata

    Database Machine to Support SLOs forMajor Workloads

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    43/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 43

    Model can be used to find optimum physical and

    virtual configurations capable of supporting SLOs

    for growing workloads

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    44/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 44

    Predicted Impact of the New VM

    Adding new VM increases contention for resources Performance prediction results illustrate the impact of adding

    VMs to the same host server and help to generate proactive

    capacity planning recommendations

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    45/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 45

    Workloads

    SLOs, SLAs

    Action

    Predictive Analytics as a Base for Defining

    Optimum Rules for Resource Management

    Modeling

    Optimization

    System

    CRM

    HR

    ETL

    Sales

    MKT

    SLMDecisions

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    46/48

  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    47/48

    Boris Zibitsker, BEZ Systems Predictive Analytics for IT 47

    Thank you!

    Contact Info rmation

    [email protected]

    bezsys.blogspot.com

    mailto:[email protected]://www.bez.com/http://www.bez.com/mailto:[email protected]
  • 8/12/2019 Presentation - Capacity Management for Oracle ExadataDatabase Machine v2

    48/48

    References

    1. E. Lazowska and others Quantative Systems Performance

    2. L. Kleinrock, Queueing Systems

    3. B. Zibitsker, A. Lupersolsky , IOUG 2009, Modeling and Optimization in Virtualized Multi-tier Distributed

    Environment

    4. B. Zibitsker, IOUG 2008. Reducing Risk of Surprises in Changing Multi-tier Distributed Oracle RAC

    Environment

    5. B. Zibitsker, DAMA 2007, Enterprise Data Management and Optimization6. B. Zibitsker, CMG 2008, 2009 Hands on Workshop on Performance Prediction for Virtualized Multi-tier

    Distributed Environments

    7. J. Buzen, B. Zibitsker, CMG 2006, Challenges of Performance Prediction in Multi-tier Parallel

    Processing Environments

    8. B, Zibitsker, G. Sigalov, A. Lupersolsky Modeling and Proactive Performance Management of Multi-tier

    Distributed Environments, International conference Mathematical methods for analysis and

    optimization of information and telecommunication networks"

    9. B. Zibitsker, C. Garry, CMG 2009, "Capacity Management Challenges for the Oracle DatabaseMachine: Exadata v2

    10.Oracle Enterprise Manager Grid Control documentation library at:

    http://www.oracle.com/technology/documentation/oem.html

    http://www.oracle.com/technology/documentation/oem.htmlhttp://www.oracle.com/technology/documentation/oem.html