oracle exadata architecture v9

Upload: kinankazuki104

Post on 03-Jun-2018

249 views

Category:

Documents


3 download

TRANSCRIPT

  • 8/12/2019 Oracle Exadata Architecture V9

    1/28

    Architecture: Oracle Exadata vs. Teradata

  • 8/12/2019 Oracle Exadata Architecture V9

    2/28

    3/6/2014 Teradata Confidential Information 2

    Topics

    Exadata Overview> Exadata Storage Server

    > HP Oracle Database Machine

    Architectural Assessment

    Pricing Review

    Summary

  • 8/12/2019 Oracle Exadata Architecture V9

    3/28

    3/6/2014 Teradata Confidential Information 3

    Oracle Exadata Storage Server

    Specialized Storage for Oracle databases

    Oracle Exadata software + HP hardware

    Addresses Oracles Disk IO Bottleneck> More, small storage units

    > Infiniband for higher bandwidth

    > Filter rows & columns in storage server

    Based on HP DL180 G5 Server> 2 Intel quad-core Xeon processors

    > 8 GB RAM

    > Array of 12 SAS or SATA disks 450 GB SAS disks

    1TB SATA disks

    > Infiniband links to Oracle DBMS> Software pre-installed Exadata software

    HP Management software

    Oracle Enterprise Linux

    Building block of Storage server cluster> 18 Exadata servers per rack

    > Multiple Racks possible

  • 8/12/2019 Oracle Exadata Architecture V9

    4/28

    3/6/2014 Teradata Confidential Information 4

    Oracle Exadata Storage Architecture

    OracleDBMS

    Infiniband

    ExadataSoftware

    Disk array

    ExadataSoftware

    Disk array

    ExadataSoftware

    Disk array

    OracleDBMS

    OracleDBMS

    OracleDBMS

    Oracle Exadata

    OracleDBMS

    OracleDBMS

    OracleDBMS

    OracleDBMS

    ArrayController

    ArrayController

    Oracle before Exadata

    Disk array Disk arrayDisk array

    Fibre Channel

  • 8/12/2019 Oracle Exadata Architecture V9

    5/28

    3/6/2014 Teradata Confidential Information 5

    Oracle Exadata 2 layer query processing

    Oracle Database layer> Aggregation

    > Sort, group by

    > Data redistribution

    > Complex joins non-partition-wise

    > Functions, UDFs, stored procs

    > SQL OLAP extensions

    > All other SQL processing

    Exadata layer

    > Projection select list columns

    > Restriction where clause predicates

    > Simple joins fact to dimensions

    ExadataSoftware

    Exadata Server

    960 MBps

    OracleDBMS

    Software

    Oracle Server

    Infiniband

    ExadataLayer

    OracleDBMSLay

    er

  • 8/12/2019 Oracle Exadata Architecture V9

    6/28

    3/6/2014 Teradata Confidential Information 6

    HP Oracle Database Machine

    Preconfigured system 8 DL360 Oracle Database servers

    > 2 quad-core Intel Xeon, 32GB RAM> Oracle Enterprise Linux> Oracle RAC

    14 Exadata Storage Cells (SAS orSATA)> Up to 20 TB uncompressed user data

    (SAS)> Up to 46 TB uncompressed user data

    (SATA)

    4 InfiniBand switches 1 Gigabit Ethernet switch Growth to multiple racks

  • 8/12/2019 Oracle Exadata Architecture V9

    7/28

    3/6/2014 Teradata Confidential Information 7

    Oracle Exadata Product Specifications

    Storage Type Storage

    Capacity

    User Data

    Volume*

    Data

    Bandwidth

    Oracle ExadataStorage Server

    450 GB SAS 5.4 TB 1-1.5 TB .96 GBps

    1 TB SATA 12 TB 2.7-3.5 TB .75 GBps

    HP Oracle

    DatabaseMachine

    450 GB SAS 75 TB 15-20 TB 13.4 GBps

    1 TB SATA 168 TB 37-46 TB 10.5 GBps

    * Varies from 3.5 disk to data ratio, to a more typical 4.5-5 ratio

  • 8/12/2019 Oracle Exadata Architecture V9

    8/28

    3/6/2014 Teradata Confidential Information 8

    Architecture Assessment of Oracle Exadata

    Compared to Teradata: Exadata does not put intelligence closer to the storage

    Exadata is not shared nothing

    Exadata does not enable high concurrency

    Exadata does not support active data warehousing Exadata does not provide superior query performance

    Exadata is not green

    Exadata is not simple

    > Trying to solve a software problem with hardware Exadata is not less expensive

  • 8/12/2019 Oracle Exadata Architecture V9

    9/283/6/2014 Teradata Confidential Information 9

    Exadata Intelligence not closer to storage

    Both read data from disk intoIntel Xeon processors

    Oracle divides query processingbetween Exadata & DB layers> Exadata does simple filtering

    > DBMS does everything else Teradata performs all query

    processing in single layer

    Oracle interconnect moves dataand Interprocess Communication

    Teradata BYNET only handlesIPC traffic

    Teradata puts all intelligence asclose to the data as Exadataintelligent storage layer

    Arraycontroller

    CPUCPU

    Teradata Node

    2.8 GBps

    375 MBps

    Arraycontroller

    CPUCPU

    Exadata Server

    960 MBps

    ASM

    DBMS

    Oracle Server

    2 GBps

    RAC

  • 8/12/2019 Oracle Exadata Architecture V9

    10/283/6/2014 Teradata Confidential Information 10

    Exadata not shared nothing

    Exadata servers each have their own disks Each partition striped across all Exadata disks

    SAME policy implemented by Oracle ASM by default

    Software mirrored via ASM failure groups

    Supports parallel IO across shared storage

    ASM

    DB

    ASM

    DB

    Exadata Exadata Exadata

    . . .

    1

    2

    3

    4

    Partitions

  • 8/12/2019 Oracle Exadata Architecture V9

    11/283/6/2014 Teradata Confidential Information 11

    Teradata shared nothing

    Each partition on separate disks not shared

    Each partition owned by separate Amp

    RAID 1 hardware mirroring

    TeradataTeradataTeradata

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

    Partitions

  • 8/12/2019 Oracle Exadata Architecture V9

    12/283/6/2014 Teradata Confidential Information 12

    Query processing Exadata vs. Teradata

    Show total sales by store for last year

    Each worker reads all Exadata disks

    Yields intra-query disk contention

    Each AMP reads its own disks

    No intra-query disk contention

    AMP1

    AMP2

    AMP3

    AMP4

    1 2 3 4

    Worker1

    Worker2

    Worker3

    Worker4

    Exadata Exadata Exadata

    1

    23

    4

  • 8/12/2019 Oracle Exadata Architecture V9

    13/283/6/2014 Teradata Confidential Information 13

    Exadata does not enable high concurrency

    Exadata reads large blocks 4 or 8 MB

    Each parallel worker reads from all Exadata disks At 80 MBps 10 concurrent IOs possible

    Intra-query and inter-query contention lower throughput

    Under high concurrency, DB parallel query worker processesqueue for IO

    ASM

    DB

    ASM

    DB

    Exadata Exadata Exadata

    . . .

    IO

    Queue

    ExadataDisk

    ParallelQuery 1

    ParallelQuery 2

    ParallelQuery 3

    ParallelQuery 4

  • 8/12/2019 Oracle Exadata Architecture V9

    14/283/6/2014 Teradata Confidential Information 14

    Exadata does not support active updates

    Oracle uses Multiple Version Concurrency Control (MVCC)> Ensures consistency of data in an active update environment> SCN checking (for correct data version) done in DB layer> SCN stored on data block

    Exadata has two modes> Filter rows and columns and return result set> Passthrough unfiltered data blocks

    Exadata returns unfiltered data blocks (passthrough)

    not filtered query results.

    ASM

    DB

    ASM

    DB

    Exadata Exadata Exadata

    . . .

    12033

    12035

    12029

    12033

    12033

    12035

    12029

    12033

    12033

    Select

    (SCN 12033)

    RollbackSegment

    Scan path

    Data blocks

  • 8/12/2019 Oracle Exadata Architecture V9

    15/283/6/2014 Teradata Confidential Information 15

    ExadataSoftware

    Exadata Server

    960 MBps

    OracleDBMS

    Software

    Oracle Server

    2 GBps

    Exadata Performance Simple Query

    1. Query coordinator dispatches parallel Exadataquery plan

    2. Each parallel slave requests exadata to readsales partitions from disk

    3. Exadata filters columns & rows> Store, date, amount

    > For current year

    4. Exadata sends matching tuples to parallelslaves

    5. Parallel slaves perform local group by and sum

    6. Parallel slaves send sums by store to querycoordinator

    7. Query coordinator performs global merge andreturns result set to requester

    Query: Show total sales by store for YTD

    1

    2

    3 4

    5

    6

    7

  • 8/12/2019 Oracle Exadata Architecture V9

    16/283/6/2014 Teradata Confidential Information 16

    ExadataSoftware

    Exadata Server

    960 MBps

    OracleDBMS

    Software

    Oracle Server

    2 GBps

    Exadata Performance Complex Query

    1. Query coordinator dispatches parallel Exadata queryplan

    2. Each parallel slave requests exadata to read salesand customer data from disk

    3. Exadata filters columns & rows

    > Store, date, amount for current year> Customer demographics

    4. Exadata returns matching tuples to parallel slaves

    5. Parallel slaves redistribute customer data for join

    6. Parallel slaves perform join

    7. Parallel slaves perform local group by and sum8. Parallel slaves send sums by store to query

    coordinator

    9. Query coordinator performs global merge andreturns result set to requester

    Query: Show sales by store by customer age, sex, income

    1

    2

    3 4

    56

    7

    98

  • 8/12/2019 Oracle Exadata Architecture V9

    17/283/6/2014 Teradata Confidential Information 17

    Exadata Query Performance

    Simple Query

    ExadataSoftware

    Exadata Server

    960 MBps

    OracleDBMS

    Software

    Oracle Server

    2 GBps

    1

    2

    3 4

    5

    6

    7

    ExadataSoftware

    Exadata Server

    960 MBps

    OracleDBMS

    Software

    Oracle Server

    2 GBps

    1

    2

    3 4

    5 6

    7

    98

    Complex Query

    As Query Complexity grows

    > Exadata portion gets smaller

    > Benefit of Exadata lessens

    > Standard Oracle DW issues

    grow Exadata biggest impact

    > Simpler star schemas

    > Data mart applications

    Exadata smallest impact

    > Complex query

    > Complex schema

    > Mixed workload

    > Ad hoc analysis

  • 8/12/2019 Oracle Exadata Architecture V9

    18/283/6/2014 Teradata Confidential Information 18

    Oracle Exadata Contribution to Performance

    Query Complexity

    ResourceAlloca

    tion(%)

    25

    50

    75

    100

    Oracle DBMS

    Exadata Cells

    Low High

  • 8/12/2019 Oracle Exadata Architecture V9

    19/283/6/2014 Teradata Confidential Information 19

    WorkloadComplexity

    OPERATIONALIZINGWHAT IShappening?

    ACTIVATINGMAKE it happen!

    ANALYZINGWHY

    did it happen?

    REPORTINGWHAT

    happened?

    PREDICTINGWHAT WILL

    happen?

    Query Performance Summary

    Data Sophistication

    Exadata Teradata

    Very good Very good

    Exadata Teradata

    Fair Very good

    Exadata Teradata

    Poor Excellent

    Exadata Teradata

    Poor Excellent

    Exadata Teradata

    Poor Excellent

  • 8/12/2019 Oracle Exadata Architecture V9

    20/283/6/2014 Teradata Confidential Information 20

    Teradata vs. Oracle Exadata per rack

    Teradata DWA2550

    HP Oracle DBMachine

    20 TB Max. User Data 15 TB

    144 x 450 GB@ 15K

    Active Disks168 x 450 GB

    @ 15K

    144 AMPs DoP Dynamic

    32 Xeon64-bit @2.33 GHz

    DatabaseProcessor Cores

    176 Xeon (112/64)64-bit @2.66 GHz

    128 GB Memory 368 (112/256) GB

    Linux 64-bit OS OEL 64-bit

    9.6 GB/s Peak Scan Rate 13.6 GB/s

    HW RAID 1 Data Protection SW Mirroring

    Oracle Exadata throwing hardware at a software problem

  • 8/12/2019 Oracle Exadata Architecture V9

    21/283/6/2014 Teradata Confidential Information 21

    Exadata Pricing _ Oracles Claims

    US$4000 per TBfor hardware

    < US$14000 perTB for combinedhardware and

    software HP Oracle

    DatabaseMachine pricingof $1.68Msoftware, $650khardware.

    Use your existing

    database licenses

  • 8/12/2019 Oracle Exadata Architecture V9

    22/283/6/2014 Teradata Confidential Information 22

    Reality: Including Oracle Database Licenses

    Software List Price/CPU Subtotal

    Oracle 11 Enter rise Edition 47 500 1 520 000

    Real A lication Clusters $23,000 $736,000

    Partitionin $11,500 $368,000

    Com ression 11 500 368 000

    Enterprise Manager Diagnostics

    Pack

    $3,500 $112,000

    Enterprise Manager Tuning Pack $3,500 $112,000

    TOTAL $100,500 $3,216,000

    Component Price

    Hardware $650,000

    Exadata Software $1,680,000

    Oracle RDBMS Software $3,216,000

    $5,546,000

    Oracleleftout

    60%of the

    listprice!

  • 8/12/2019 Oracle Exadata Architecture V9

    23/283/6/2014 Teradata Confidential Information 23

    Oracle Database Machine vs. Teradata Family

    Teradata DW Appliance less expensive than Oracle DBM w SAS disks> Fully loaded 4x advantage> Exadata only 2x advantage

    Teradata Extreme DWA less expensive than Oracle DBM w SATA disks

    Teradata ADW less expensive than Oracle DBM fully loaded

    SpinningDisk

    UserData

    ExadataPrice

    TotalPrice

    Oracle DBM w SAS disks

    (450 GB @ 15K RPM)75 TB 15 TB $2.3M $5.5M

    Oracle DBM w SATA disks

    (1 TB @ 7200 RPM)168 TB 37 TB $2.3M $5.5M

    Teradata ADW

    (300 GB @ 15K RPM)44 TB 28 TB -- $4.2M

    Teradata DW Appliance

    (450 GB @ 15K RPM)43 TB 20 TB -- $1.3M

    Teradata Extreme DWAppliance (1 TB @ 7200 RPM) 124 TB 50 TB -- $0.8M

  • 8/12/2019 Oracle Exadata Architecture V9

    24/283/6/2014 Teradata Confidential Information 24

    Oracle Database Machine vs. Teradata Family

    Teradata DW Appliance> less expensive than Oracle DBM fully loaded (SAS or SATA)> Less expensive than Oracle DBM w SAS (Exadata only)

    Teradata ADW less expensive than Oracle DBM w SAS (fully loaded)

    Teradata Extreme DW Appliance cheaper than Oracle w SATA

    TeradataADW

    TeradataDW

    Appliance

    TeradataExtreme

    DWAppliance

    OracleDBM

    w SATA

    OracleDBM

    w SAS

    $100K

    $200K

    $300K

    $400K

    Exadata onlyOracle DBMS

    Teradata DBMS

    $150K

    $65K

    $16.5K

    $150K

    $370K

    $155K

    $63KPriceper

    TBofUserData

  • 8/12/2019 Oracle Exadata Architecture V9

    25/283/6/2014 Teradata Confidential Information 25

    Summary

    Oracle Data Base Machine :> Is not ground breaking

    > Is not cheap

    > Is not an appliance - still multiple support points

    > Is not an EDW solution EDW requires more than fastscans

    > Does not solve most of Oracle's DW performance problems

    > HP Oracle DMB is not a balanced system

    > Is unnecessarily complex

    So what does it do?> Address a Oracle configuration problem

    > Emphasize the OOWI reference arch

  • 8/12/2019 Oracle Exadata Architecture V9

    26/283/6/2014 Teradata Confidential Information 26

    Thank You!

  • 8/12/2019 Oracle Exadata Architecture V9

    27/283/6/2014 Teradata Confidential Information 27

    TeradataDBMS

    Software

    Teradata Node

    2.8 GBps

    375 MBps

    Teradata Performance Simple Query

    1. PE dispatches parallel query plan to AMPs

    2. Each AMP to read partition data from disk

    3. AMPs filter columns & rows

    > Store, date, amount

    > For current year

    4. AMPs perform local group by and sum

    5. AMPs redist local groups for global merge

    6. AMPs perform global merge and returns result

    set to PE7. PE returns result set to requester

    Query: Show total sales by store for YTD

    1

    2

    3

    4

    5

    6

    7

  • 8/12/2019 Oracle Exadata Architecture V9

    28/28

    TeradataDBMS

    Software

    Teradata Node

    2.8 GBps

    375 MBps

    Teradata Performance Complex Query

    1. PE dispatches parallel query plan to AMPs

    2. Each AMP reads sales and customer data fromdisk

    3. AMPs filters columns & rows

    > Store, date, amount for current year> Customer demographics

    4. AMPs redistribute customer data for join

    5. AMPs perform join

    6. AMPs perform local group by and sum

    7. AMPs redist local groups for global merge8. AMPs perform global merge and returns result

    set to PE

    9. PE returns result set to requester

    Query: Show sales by store by customer age, sex, income

    1

    2

    3

    4

    5

    6

    7 9

    8