Professional Documents
Culture Documents
Teradata Tools
Teradata Tools
Architecture, Technology, Scalabilty, Performance and Vision for Active Enterprise Data Warehousing
Dr. Barbara Schulmeister Teradata a Division of NCR
Barbara.Schulmeister@ncr.com
28. 6. 2005
Agenda
History Definitions Hardware Architecture Fault Tolerance and High Availability Coexistence Operational System Tools and Utilities Data Distribution SQL Parser Active Data Warehouse Scalability
Born to be parallel!
DBC Model 1: First MPP System! Product of the Year Forbes DBC Model 3 Fastest Growing Small Company INC Magazine Fastest Growing Electronic Company Electronic Business Leader in Commercial Parallel Processing Gartner Group
1979...
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994...
DBC Model 4
3+ TB System!
more
...1995
1996
1997
...
...only NCRs Teradata V2 RDBMS has proven it can scale Gartner Group
more
...1998
1999
2000
2001
2002
2003
2004
2005
IT Award of Excellence
V2R5 Teradata
V2R6 Teradata
Linux
TDWI Solution Provider Best Practices in Data Warehousing TDWI Leadership in Data Warehousing Award DM Review World-Class Solution Award for business Intelligence IT Times Award DM Review 100 Award DM Review Readership Award Intelligent Enterprise Real Ware Award
Only viable for low volume access Meta data issues Network bandwidth and join complexity issues Workload typically placed on workstation
Business Enterprise view challenging Redundant data costs High DBA and operational costs Data latency
DSS
NCR Server
Provide customers with growth opportunities and investment protection
> Coexistence is enabled across five generations
NCR 5400E & 5400H Servers NCR 4980 & 5380 Servers NCR 4950 & 5350 Servers NCR 4900 & 5300 Servers NCR 485X & 525X Servers BYNET V2 / V3
> 1 - 4 nodes > BYNET V2 > ESCON & FICON for 3 and 4 node configurations > Field Upgradeable to 5400H
Up to 4 nodes within each cabinet Server Management
1 3
nd 2 2nd
Node Node st Node 1 1st Node 4th Node 3rd Node 3GSM
1 3
1 3
1 3
1 3
1 3
1 3
1 3
1 3
1 3
1 3
1 3
1 3
1 3
Multi-Generation Coexistence
> Investment protection
2500
54000
2000
1500
1000
Rock 90nm
500
FUJITSU Primepower
HP HP9000
HP Integrity
HP Proliant
IBM pSeries
NCR Teradata
SUN Sunfire 40
PRODUCT
43
45
46
29
45
54
The Product category (which was called Technology in previous ASEM updates) focuses on the performance and reliability/availability aspects of each platform. In this category Teradata received a very strong 93.5% of total possible points and leads the IBM pSeries with 74.35% by 44 points or 19%.
Source Gartner 2004 ASEM Report
> Dual Quad Fibre Channel Controllers per array for performance and availability > Typical configuration is 4 NCR 5400 Server nodes per 3 6842 arrays
1.2 Terabytes of database space per node (RAID 1)
> Supports RAID 1 and RAID 5 > Support for MP-RAS and Microsoft Windows Server 2003 environments
EMC Model Disks Teradata Use RAID Options Operating Environment Maximum Teradata disks
DMX 1000 M2 73GB 15K RPM MPP: supports 1 or 2 nodes per cabinet RAID -1 Only MP-RAS and Windows 96
DMX 2000 M2 73GB 15K RPM MPP: supports 2, 3, or 4 nodes per cabinet RAID-1 Only MP-RAS and Windows 192
All system configurations discussed in this presentation are based on balanced or supersaturated compute nodes.
Optimum
e im eT ns po es yR er Qu
th id dw n Ba
Performance more than adequate: Add more data to all nodes Query Response Time
Query Response Time Increases because you didnt add more compute power to support the additional raw data volume.
Query Response Time Remains Constant because you add proportionally more raw data volume as compute power.
Raw Data Volume adequate: Upgrade to faster CPUs Query Response Time
SYSTEM with Current Nodes SYSTEM with More or Faster Nodes
Query Response Time Decreases because you didnt add more raw data volume to offset the increase in compute power. Scale vertically with Teradata by increasing compute power.
Scale to Target query performance and data volume by increasing compute power and adding storage.
Improve query performance and adjust data volume by reducing storage per node and adding more nodes.
BYNET Fabrics
CPU uses independent direct I/O path to Disk All memory accesses are local
Disk Storage Disk Storage Disk Storage Disk Storage Disk Storage Disk Storage Disk Storage
Storage
SMP Node1
CPU1 CPU2
SMP Node2
CPU1 CPU2
SMP Node3
CPU1 CPU2
SMP Node4
CPU1 CPU2
Memory
Memory
Memory
Memory
Server Management
> One console to view the entire system
Server Management
PE1 PE2
AMP5
Access Module Processor VPROCS
VPROCS
LAN Gateway
Communication Interfaces
AMP4
Disk Array
Channel Gateway
PEs recieve the queries and figure out the query plan AMPs interact with the disk arrays and process the data
PE VProc
Parser Optimizer Session Control
VAMP
Relational Database Management File System / Data Management
Dispatcher
Broadcast Messaging
Node Node
The Teradata Optimizer chooses between Point-to-Point and Broadcast Messaging to select the most effective communication.
Bandwidth scales linearly to 1,024 nodes Redundant, fault tolerant network Guaranteed message delivery
Work migrates to the remaining nodes in the cliques. System performance degradation up to 33%.
Traditional Configuration
Large Cliques
Double the number of nodes in a clique up to 8. Work distributed across a greater number of nodes. Minimize system performance impacts may not be noticeable to end-users.
Node
Node
Node
Node
Node
Node
Node
Node
Disk Array
Disk Array
Disk Array
Disk Array
Disk Array
Disk Array
X
Node Disk Array
Node
Node
Hot Standby
Disk Array
Disk Array
X
Node
Node
Node
Node
Node
Node
Node
Hot Standby
Disk Array Disk Array Disk Array Disk Array Disk Array Disk Array
High Availability
Case
Power Failure Node Failure Bynet failure Disc failure More than one Disc Failure Clique Failure
Hardware
UPS (redundant), Dual AC
Teradata
VPROC Migration (VAMP, PE) Redundant BYNET RAID-1/-5/-S in Disc Subsystem Fallback-Option Fallback-Option
BYNET
Server Nodes
Coexistence Considerations
Generation x VAMPS
Generation x VAMPS
AMP AMP AMP AMP AMP AMP AMP AMP AMP AMP
Generation x VAMPS
AMP AMP AMP AMP AMP AMP AMP AMP AMP AMP AMP AMP AMP AMP AMP AMP
AMP AMP
AMP AMP
AMP AMP
AMP AMP
AMP AMP
Performance Factor
1x
1.5x
2.0x
VAMPs manage the same amount of data Coexistence enables the faster nodes to be realized by running more VAMPs per node
SMP Node1
AMP AMP AMP
SMP Node2
AMP AMP AMP
SMP Node3
AMP AMP AMP
SMP Node4
AMP AMP AMP
SMP Node1
AMP AMP AMP
SMP Node2
AMP AMP AMP
SMP Node3
AMP AMP AMP
SMP Node4
AMP AMP AMP
Server Management
Generation A
Generation B
Generation C
Generation D
Generation E
485x/525x
4900/5300
4950/5350
4980/5380
5400
V2R6 on WS03 (2Q 2005) V2R5.0.3, V2R5.1.X, V2R6 on W2K (2Q 2005) Database > Teradata V2R6 > Support one Release Back V2R5.1.X (Current Exception in Place V2R5.0.3) Unix > MP-RAS 3.03 required for Teradata Database on 5400 > MP-RAS 3.02 still supported on previous server generations Microsoft Windows > Microsoft Windows Server 2003 recommended for new and expanding 5400 motions > Microsoft Windows 2000 supported in 2Q 2005
Partial Covering Join Index Global Index Sparse Index Join Index Extensions ODS Workload Optimization Stored Procedures Enhancements Enhancements to Triggers Extra FK-PK Joins in Join Index UDFs for XML Processing etc.
Security enhancements (Encryption) DBQL enhancements Database Object Level Use Count ROLES enhancements Priority Scheduler enhancements TDQM enhancements No Auto Restart After Disk-Array Power Failure Cancel Rollback Incompatible Package Warning Disk I/O Integrity Check
Data Freshness
Cylinder Read Partitioned Tables (PPI) Value List Compression 2000 Columns, 64 Columns per Index Identity Column Enhancement to Identity Column UTF16 Support PPI Dynamic Partition Elimination Large Objects (LOBs)
Continuous Continuous Update Update Performance Performance & & Manageability Manageability Faster Faster Join Join Index Index Update Update Join Join Update Update Performance Performance Bulk Bulk Update Update Performance Performance Teradata Teradata Warehouse Warehouse Builder Builder Full Full Functionality Functionality & & Platform Platform Support Support UDFs for Data Transformation and Scoring
Improve Primary Index Operations\ Improved IN-list processing External Stored Procedures Trigger calling a Stored Procedure Stored Procedure Internals Enhancements Queue tables
Data Freshness
Stored Procedure LOB support External Table Function Partition level BAR Eliminate indexed row IDs (PPI) PPI Join performance improvement DBS Information consolidation
Replication Services Array support Priority Scheduler enhancements Reduce restart time
Database Management
Teradata Manager Teradata Dynamic Query Manager Teradata System Emulation Tool Teradata Visual Explain Teradata Index Wizard Teradata Statistics Wizard
Metadata
Teradata Meta Data Services
Mainframe Connectivity
Mainframe Channel Connect TS/API, CICS, HUTCNS & IMS/DC
Any Query, Any Time
Teradata utilities are fully parallel. Teradata utilities have checkpoint restart capability. Data loads directly from the source into the database.
> > > > No No No No manual data partitioning. file splitting. intermediary file transfers. separate data conversion step.
Parallel In Parallel Out
Teradata Warehouse
Two Basic Software Architecture Models Task Centric and Data Centric
Request
Request
Request
Request
Task
Shared Memory
Task
Parallel Optimizer
Parallel Unit
DATA
DATA
Data Data
Data
Uniform and shared access to all platform resources (disk, etc) is REQUIRED
Processor
D
Memory
Disk
AMP 1
P
AMP 2
P
AMP 3
P
AMP 4
P
25 21 17 13 9 5 1
26 22 18 14 10 6 2
27 23 19 15 11 7 3
28 24 20 16 12 8 4
Table A
Table B
Table C
Each virtual AMP has rows from every table Each virtual AMP works independently on its rows Goal: Database rows are equally distributed across multiple tables
Row Hash (32 bits) Current configuration Primary Current configuration Fallback Reconfiguration Primary Reconfiguration Fallback
Hash Map
Node1
Node2
Node3
Node4
Node1
Node2
Node3
Node4
Node1
Node2
Node3
Node4
Node1
Node2
Node3
Node4
Teradata Hashing
Table ORDER
O rder Num ber C u s to m e r Num ber O rder D a te O rder S tatus
PK UPI 7325 7324 7415 7103 7225 7384 7402 7188 7202 2 3 1 1 2 1 3 1 2 4/13 4/13 4/13 4/10 4/15 4/12 4/16 4/13 4/09 O O C O C C C C C
(Hexadecimal)
HASH MAP 6 07 07 07 07 07 07 7 08 08 08 08 08 08 8 01 01 01 01 01 01 9 02 02 02 02 02 02 A 03 03 03 03 03 03 B 04 04 04 04 04 04 C 05 05 05 05 05 05 D 0 0 0 0 0 0
0 01 01 01 01 01 01
1 02 02 02 02 02 02
2 03 03 03 03 03 03
3 04 04 04 04 04 04
4 05 05 05 05 05 05
5 06 06 06 06 06 06
AMP 1
AMP 2
AMP 3
AMP x
DISTRIBUTION
Optimize parallel processing: choose a column that provides good distribution
Volatility
Reduce maintenance resource overhead (I/O): choose a column with stable data values
48 bit table ID
Index value
Node1
Node2
Node3
Node4
Node1
Node2
Node3
Node4
Node1
Node2
Node3
Node4
Node1
Node2
Node3
Node4
Cached?
Syntaxer
DBase, AccRights TVM, TVFields Indexes
DD
Resolver Security
Serial steps Parallel steps Individual and common steps (MSR) Additional: Triggers, check constraints, references, foreign keys, join indexes collected statistics or dynamic sampling
Statistics
Optimizer
Costs
Statistics Summary
Collect statistics on all non-unique indexes UPI of any table with less than x rows per AMP (dependent on available number of AMPs) All indexes of a join index any non-indexed column used for join constraints indexes of global temporary tables Collected statistics are not automatically updated by the system Refresh statistics when 5-10% of the table rows have changed
Tactical (ODS)
User Profiles Front Line Services Customers - Indirectly Services: Lookups Tactical decisions Analytics (e.g. scoring) Access Profile: Continuous updates Frequent lookups Data Model: Current state data Recent history Integrated business areas
Strategic (EDW)
User Profiles Back Office Services Management Trading Partners Services: Strategic decisions Analytics (e.g. scoring) Access Profile: Bulk Inserts Some Updates Frequent complex analytics Data Model: Periodic state data Deep history Enterprise integrated view
Workload Continuum
OLTP1
OLTPi
OLTPn
ODS1
ODS2
Transactional Repositories
Tactical (ODS)
User Profiles Front Line Services Customers - Indirectly Services: Lookups Tactical decisions Analytics (e.g. scoring) Access Profile: Continuous updates Frequent lookups Data Model: Current state data Recent history Integrated business areas
Strategic (EDW)
User Profiles Back Office Services Management Trading Partners Services: Strategic decisions Analytics (e.g. scoring) Access Profile: Bulk Inserts Some Updates Frequent complex analytics Data Model: Periodic state data Deep history Enterprise integrated view
Workload Continuum
OLTP1
OLTPi
OLTPn
Transactional Repositories
Workload Complexity
Query complexity grows Workload mixture grows Data volume grows Schema complexity grows Simultaneous Workloads: Depth of history grows Strategic, tactical, Number of users grows loading Expectations grow
Increasing depth and breadth of users and queries ANALYZING WHY did it happen? REPORTING WHAT happened?
Increase in ad hoc analysis Primarily batch and some ad hoc reports
Data Sophistication
Single View of the Business Better, Faster Decisions Drive Business Growth
Workload Complexity
Query complexity grows Workload mixture grows Data volume grows Schema complexity grows Depth of history grows Number of users grows Expectations grow
Execute
Continuous update and time-sensitive queries become important
Optimize
Batch
Measure
Data Sophistication
Single View of the Business Better, Faster Decisions Drive Business Growth
Workload Complexity
Query complexity grows Workload mixture grows Data volume grows Schema complexity grows Depth of history grows Number of users grows Expectations grow
Data Sophistication
Single View of the Business Better, Faster Decisions Drive Business Growth
Teradata systems can address this need through a combination of technologies, such as:
> Partitioned primary index (PPI). > Multi-value compression. > Priority scheduler.
Pre-Execution
Manage the level of resources allocated to different priorities of executing work Priority Scheduler
ADW
Post-Execution
Indexes
PI (UPI and NUPI) SI (USI and NUSI) Join Index single table index multi table index aggregated index sparse index (where clause used) partial covering global Materialized Views (join index)
An Integrated, Centralized Data Warehouse Solution Database Must Scale in Every Dimension
Data Volume (Raw, User Data) Mixed Workload Query Concurrency
Data Freshness
Query Complexity
Schema Sophistication
An Integrated, Centralized Data Warehouse Solution Database Must Scale in Every Dimension
Data Volume (Raw, User Data) Mixed Workload Query Concurrency
Data Freshness
Query Complexity
Schema Sophistication
Mixed Workload
Customers Need to Evaluate Real Life Workloads
Query Concurrency
Good Example? TPC-H Benchmark
Data Freshness
Query Complexity
Query Freedom
Schema Sophistication
Companies generating >80% of the industry revenue utilize Teradata Data Warehousing
Global presence
> Over 100 countries
Global presence
> Over 100 countries
www.teradata.com
Data Volume (Raw, User Data) Mixed Workload Query Concurrency
Data Freshness
Query Complexity
Schema Sophistication