Partitioning with Oracle 11G

Download Report

Transcript Partitioning with Oracle 11G

Partitioning with
Oracle 11G
Bert Scalzo, Domain Expert, Oracle Solutions
[email protected]
Copyright © 2006 Quest Software
About the Author …
Domain Expert & Product Architect for Quest Software
Oracle Background:
•
Worked with Oracle databases for over two decades (starting with version 4)
•
Work history includes time at both “Oracle Education” and “Oracle Consulting”
Academic Background:
•
Several Oracle Masters certifications
•
BS, MS and PhD in Computer Science
•
MBA (general business)
•
Several insurance industry designations
Key Interests:
•
Data Modeling
•
Database Benchmarking
•
Database Tuning & Optimization
•
"Star Schema" Data Warehouses
•
Oracle on Linux – and specifically: RAC on Linux
Articles for:
•
Oracle’s Technology Network (OTN)
•
Oracle Magazine,
•
Oracle Informant
•
PC Week (eWeek)
This presentation draws
heavily on these areas
Articles for:
•
Dell Power Solutions
Magazine
•
The Linux Journal
•
www.linux.com
•
www.orafaq.com
1
Books by Author …
Coming in 2008 …
2
Agenda
• Partitioning Benefits
• Partitioning History
• Partitioning Options
• Partitioning Advisor (if you’re licensed)
• Typical Data Warehousing Environment
• TPC-H “Data Warehouse” Benchmark
• Results TPC-H with Various Partition Strategies
• What about OLTP Environments and the TPC-C/E
• Lessons Learned (and their relevance/application)
• Questions & Answers
3
Partitioning Benefits: Facts
• Manageability
– Classic “Divide & Conquer” technique
– More granular storage allocation options
– Keeps otherwise time consumptive options viable
• Availability
– More granular online/offline options
– More granular rebuild/reorganization options
– More granular object level backup/restore options
• Capacity Management
– Enables a “Tiered Storage Architecture” approach
– More granular storage cost management decision points
• Performance
– Partition Pruning
– Partition-Wise Joins
4
Partitioning Benefits: Opinion (Mine)
•
•
•
•
Manageability
Availability
Capacity Management
Performance
40%
20%
20%
20%
Why to
Partition
• Don’t over-sell/over-expect the performance aspect
• Need to experiment for best approach for a database
• Better to take longer at the start to get right, because
very often it’s far too expensive to change afterwards
• Examples demonstrate very positive performance,
but better to be conservative and error on the side of
caution – then be very pleasantly surprised…
5
Partition Pruning (Restriction Based)
• From Docs: In partition pruning, the optimizer analyzes FROM and
WHERE clauses in SQL statements to eliminate unneeded partitions
when building the partition access list. This enables Oracle Database to
perform operations only on those partitions that are relevant ...
• “Divide and Conquer” for performance
– Sometimes can yield order of magnitude improvement
– But once again, best not to oversell and/or over-expect
• Some Potential Issues to be aware of:
– SQL*Plus Auto-Trace can sometimes miss partition pruning
– “Old Style” Explain Plans via simple SELECT has issues too
– Best to always use DBMS_XPLAN and/or SQL_TRACE
Note: Trace file analysis much easier these days – SQL Developer +
free Hotsos plug-in, metalink trace analysis scripts, Quest Toad DBA
6
Partition-Wise Join (Multi-Object Based)
• From Docs: Partition-wise joins reduce query response time by
minimizing the amount of data exchanged among parallel execution
servers when joins execute in parallel. This significantly reduces
response time & improves the use of both CPU & memory resources.
• Different Flavors:
–
–
–
–
–
Full – Single to Single
Full – Composite to Single
Full – Composite to Composite
Partial – Single
Partial – Composite
All of these
affect the
explain plan
• Indexing Strategy Counts
– Local Prefixed/Non-Prefixed
– Global
7
Picture Worth 1000 Words (from Docs)
Simple Mantra: Subdivide the work into equally paired
chunks, then perform all that work using many parallel
processes
Make sure not to over-allocate
CPU’s – remember there will
also be concurrent workload
8
Partitioning History (from Oracle 11G training+)
Oracle 5
Oracle 7
Before Tablespaces – we had partitions 
Partition Views – really more of a cheat 
9
Partitioning Options – Part 1
IOT’s can be partitioned as
well in later versions of Oracle,
so the basic choices are even
more complex than this…
10
Partitioning Options – Part 2
Prior to 11G:
Oracle White Paper: 2007 Partitioning in Oracle Database 11g
11
Partitioning Options – Part 3
Post 11G:
Oracle White Paper: 2007 Partitioning in Oracle Database 11g
Very
exciting
new
options…
12
Partitioning Advisor (if you’re licensed)
Advisor Central -> SQL Advisors -> SQL Access Advisor
13
Typical Data Warehouse Architecture
DM 1
OLTP
App #1
VSAM
OLTP
App #2
Sybase
ETL
ODS
Oracle
TPC-H
DM 2
Enterprise
DW
OLTP
App #3
Oracle
ET
OLTP
App #4
ISAM
Staging
Area
L
14
Typical Environments
OLTP
ODS
OLAP
DM/DW
Business
Focus
Operational
Operational
Tactical
Tactical
Tactical
Strategic
End User
Tools
Client
Server Web
Client Server
Web
Client Server
Client Server
Web
DB
Technology
Relational
Relational
Cubic
Relational
Trans Count
Large
Medium
Small
Small
Trans Size
Small
Medium
Medium
Large
Trans Time
Short
Medium
Long
Long
Size in Gigs
10 – 200
50 – 400
50 – 400
400 - 4000
Normalization
3NF
3NF
N/A
0NF
Data Modeling
Traditional
ER
Traditional ER
N/A
Dimensional
We’ll come back to this picture
15
TPC-H Benchmark
• Industry Standard “Data Warehouse” Benchmark
• URL: www.tpc.org/tpch
• Spec: http://tpc.org/tpch/spec/tpch2.7.0.pdf
• 8 Tables
• 22 Queries (answer complex business questions)
• Database scaling:
– Factor = 1, 10, 30, 100, 300, 1000, 3000, 10000, 30000, 100000
– Size GB = 1, 10, 30, 100, 300, 1000, 3000, 10000, 30000, 100000
16
Sub-Partitions
TPC-H Data Model
SF *
200,000
SF *
6,000,000
SF *
800,000
SF *
10,000
Partitions
25
5
SF *
150,000
SF *
1,500,000
17
TPC-H Permits Partitioning …
But what to do, what to do ???
18
Disclosure Reports
http://tpc.org/tpch/results/tpch_perf_results.asp
19
Disclosure Report – Lots of Info
This is where people document
exactly what advanced database
feature and storage parameters
they used – info is invaluable 
20
Disclosure Report – Appendix B
21
Sample Expensive Query
22
Example Explain Plan
23
Example Explain Plan
Explain complete.
Plan hash value: 2545634784
---------------------------------------------------------------------------------------------| Id | Operation
| Name
| Rows | Bytes |TempSpc| Cost (%CPU)| Time
|
---------------------------------------------------------------------------------------------|
0 | SELECT STATEMENT
|
| 42533 | 5648K|
|
641K (1)| 01:57:41 |
|
1 | SORT GROUP BY
|
| 42533 | 5648K|
105M|
641K (1)| 01:57:41 |
|* 2 |
HASH JOIN
|
|
715K|
92M|
|
631K (1)| 01:55:51 |
|
3 |
TABLE ACCESS FULL
| H_NATION
|
25 |
725 |
|
3
(0)| 00:00:01 |
|* 4 |
HASH JOIN
|
|
715K|
72M|
|
631K (1)| 01:55:51 |
|
5 |
TABLE ACCESS FULL
| H_SUPPLIER |
100K|
781K|
|
646
(1)| 00:00:08 |
|* 6 |
HASH JOIN
|
|
720K|
68M|
68M|
631K (1)| 01:55:44 |
|* 7 |
HASH JOIN
|
|
751K|
59M|
232M|
589K (1)| 01:48:10 |
|* 8 |
HASH JOIN
|
| 3004K|
197M| 4984K|
485K (1)| 01:28:57 |
|* 9 |
TABLE ACCESS FULL| H_PART
|
100K| 3808K|
| 11805
(1)| 00:02:10 |
| 10 |
TABLE ACCESS FULL| H_LINEITEM |
60M| 1716M|
|
342K (1)| 01:02:52 |
| 11 |
TABLE ACCESS FULL | H_ORDER
|
15M|
200M|
| 72122
(1)| 00:13:14 |
| 12 |
TABLE ACCESS FULL | H_PARTSUPP | 8000K|
122M|
| 25945
(1)| 00:04:46 |
---------------------------------------------------------------------------------------------Predicate Information (identified by operation id):
--------------------------------------------------2 - access("S_NATIONKEY"="N_NATIONKEY")
4 - access("S_SUPPKEY"="L_SUPPKEY")
6 - access("PS_SUPPKEY"="L_SUPPKEY" AND "PS_PARTKEY"="L_PARTKEY")
7 - access("O_ORDERKEY"="L_ORDERKEY")
8 - access("P_PARTKEY"="L_PARTKEY")
9 - filter("P_NAME" LIKE :SYS_B_2)
24
Method of Attack
• Since many data warehouses are utilized for data
mining, we can’t always know every possible query
likely to run – thus aggregate measure for success
• Thus we’ll compare the benchmark’s weighted
performance scores for the TPC-H using various
partitioning schemes (all within spec of course…)
• Goal will be to find the best overall partitioning …
• Then we’ll examine some specific explain plans …
25
10G Sample Test cases
•
10G Simple Approach (Just Huge Tables)
– Range: ORDER (order date)
– Hash: LINEITEM (order key)
•
10G Basic Approach (Single Level Partitions)
– Range: ORDER (order date)
– Hash: LINEITEM (order key)
– List: CUSTOMER (nation key)
– Hash: PART, SUPPLIER and PARTSUPP (part & supp keys)
•
10G Complex Approach (Composite Partitions)
– Range-Hash: ORDER (order date & cust key)
– Multi-Hash: LINEITEM (part, supp & order keys)
– List: CUSTOMER (nation key)
– Hash: PART, SUPPLIER and PARTSUPP (part & supp keys)
26
11G Sample Test cases
•
11G Simple Approach (+Interval)
– Interval-Hash: ORDER (order date & cust key)
– Multi-Hash: LINEITEM (part, supp & order keys)
– List: CUSTOMER (nation key)
– Hash: PART, SUPPLIER and PARTSUPP (part & supp keys)
•
11G Basic Approach (+Virtual)
– Interval-Hash: ORDER (virtualized order date & cust key)
– Multi-Hash: LINEITEM (part, supp & order keys)
– List: CUSTOMER (nation key)
– Hash: PART, SUPPLIER and PARTSUPP (part & supp keys)
•
11G Complex Approach (+REF)
– Interval-Hash: ORDER (virtualized order date & cust key)
– REF: LINEITEM (order key)
– List: CUSTOMER (nation key)
– Hash: PART, SUPPLIER and PARTSUPP (part & supp keys)
27
Is that It?
• No – just six very obvious high-level scenarios
• Your selections and actual mileage will vary
• Experimentation usually yields the best results
• Always trust “empirical results” over conjecture
• So improved response-time beats better explain plan
• Remember, DW’s usually have unpredictable queries
• So don’t tune for just a few queries, look for the best
overall and/or more generic performance solution 
28
Intermediate Results can be Misleading 
Power
Th-Put
450
300
400
250
350
300
200
250
150
200
150
100
100
50
50
0
bmf1
bmf2
bmf3
bmf4
bmf5
TPC-H Power score seemingly
implies that every partitioning
schema is incrementally better
0
bmf1
bmf2
bmf3
bmf4
bmf5
TPC-H Throughput score
seems to show that nonpartitioned is equal to the best
29
Final Results tell the Real Truth 
$/QpH
QpH
$14.00
350
$12.00
300
250
$10.00
200
$8.00
150
$6.00
100
$4.00
50
$2.00
0
$0.00
bmf1
bmf2
bmf3
bmf4
bmf5
TPC-H Query/Hour score
shows that some partitioning
schemes better, and some not
bmf1
bmf2
bmf3
bmf4
bmf5
TPC-H $/Query/Hour score
confirms the inverse in terms
of dollars per unit of work
30
Why such seemingly Opposite Results ???
•Run times and explain plans apply to single measurable operation
•Even aggregate & averaged run times don’t relate the entire truth!
•Need answer based upon sound mathematics (reliable & repeatable)
31
TPC-C Benchmark
• Historical Industry Standard “OLTP” Benchmark
• URL: www.tpc.org/tpchc
• Spec: http://tpc.org/tpcc/spec/tpcc_current.pdf
• Probably the most used & widely quoted benchmark
• But suffers from overly simplistic design & code logic
• Generally considered unreliable with modern RDBMS
• But still a decent rough “sounding board” for many ….
• Being replaced by the newer TPC-E (later slides)
32
TPC-C Data Model
Base Scaling Unit
# Terminals/Warehouse
(i.e. concurrent users)
Clustered
Partitions
Sub-Partitions
TPC-E Benchmark
• Emerging Industry Standard “OLTP” Benchmark
• URL: www.tpc.org/tpche
• Spec: http://tpc.org/tpce/spec/TPCE-v1.5.1.pdf
• Very new and still evolving – but highly promising
• Not too many published TPC-E results as of yet …
• Design not compromised by RDBMS features
• Much more realistic (i.e. real world) in nature 
• Nowhere near as easy as the old TPC-C test 
34
TPC-E Data Model
35
TPS is Moot, Average Response Time is King
tpmC
584
582
579
578
But wait: adding cluster &partitioning yields negative – why ???
Look to Stats Pack, AWR and ADDM Reports to investigate…
36
Single block
reads 32 ms!
DISTRICT Table
needs clustered
Suggest fix one major item
per test iteration, so made
choice to address this 1st
Single block read 32 ms 
Clustering worked, it made
SQL the #1 performance
issue – as was expected
Single block read 5 ms 
Partitioning did not shine
through just yet, possibly
skewed by the first issue
Switch MEMORY_TARGET to SGA/PGA_TARGETS
-13%
Notice that manual memory management resulted in 13% gain !!!
But wait, there’s more (isn’t that almost always the case) …
39
Now it’s all
just SQL, so
time for
SQL Tuning
Advisor
&
SQL Tuning
Sets
SQL Tuning Sets & 11G Results/Client Cache
-9%
Going to stop – Have quickly reached the “good enough” point 
41
Architecture Findings
OLTP
ODS
OLAP
DM/DW
Business
Focus
Operational
Tactical
Strategic
End User
Tools
Client
Server Web
Client Server
Web
DB
Technology
Relational
Relational
Trans Count
Large
Small
Trans Size
Small
Trans Time
Short
Size in Gigs
10 – 200
400 - 4000
Normalization
3NF
0NF
Data Modeling
Traditional
ER
Mostly
Partition
Elimination
Design to
eliminate
per object
Mostly
Partition
Wise Join
Design to
parallelize
across
objects
Large
Long
Dimensional
42
Other Interesting Findings
• 64-bit scales much more reliably than 32-bit, even when
on same hardware and using the <= 4GB memory model
• If you know the application code’s nature and it’s well
definable, manual memory management may be better
• Using manual SGA/PGA targets with floors yields much
more scalable results and also more predictable patterns
• Partitioning is not an automatic bonus, must experiment
to identify the optimal partitioning scheme per situation
• Don’t forget older technologies like clusters, they still can
add a positive to the overall equation in certain cases
• Don’t forget SQL Tuning Sets & SQL Advisor, or the
“explain plans” may not in fact be the best obtainable
• Don’t forget 11G’s Result and Client Caches
43
Questions and Answers …
Thank you
Presenter: Bert Scalzo
E-mail: [email protected]
Note: these slides should be available on Open World web site,
but I’ll also make sure to post them on my company’s web site:
www.toadworld.com/Experts/BertScalzosToadFanaticism/tabid/318/Default.aspx