Choosing Indexes For Performance

(c) 2015 Independent SAP Technical User GroupAnnual Conference, 2015
ISUG-TECH 2015 ConferenceISUG-TECH 2015 Conference
Pr esent er ’ s Sessi on Ti t l e her ePr esent er ’ s Sessi on Ti t l e her e
Pr esent er ’ s namePr esent er ’ s name

AgendaAgenda
Welcome
Speaker Introduction
Session Title (add presentation title)
Q&A

Jeff GarbusJeff Garbus
Speaker(s) should add a few introductory
commentsabout him/herself

Single – Table Optimization

Acknowledgements
Sybase Adaptive Server is a
trademark of Sybase Inc.
This presentation is
copyrighted.
This presentation is not for re-
sale
This presentation shall not be
used or modified without
express written consent of
Soaring Eagle Consulting,
Inc.
Slide 6 - 5

Examine detailed topics in query optimization
Indexes with SARGs
Improvised SARGs
Clustered vs. nonclustered Indexes
Queries with OR
Index covering
Forcing index selection
Partition support
Slide 6 -
Single - Table Optimization
Topics

Table scans
Index searches
Covered index searches
Slide 6 - 7
Adaptive Server Search Techniques
ASE uses three basic search techniques for
query resolution

Scans are expensive
Table scans may be planned even if the rest of
the query is false!
Slide 6 - 8
Table Scans
If Adaptive Server can’t resolve a query any
other way, it does a table scan
Identify table scan searches
select * from pt_tx where 1 = 2

(c) 2015 Independent SAP Technical User GroupAnnual Conference, 2015 Slide 6 - 9
Table Scans
select * from pt_tx where 1=2
SHOWPLAN OUTPUT

Table Scans
DBCC TRACEON (3604, 302,

Table Scan Output: SELECT
select *
from pt_tx
SHOWPLAN

Table Scan Output: UPDATE
update pt_tx
set id = id + 1
SHOWPLAN OUTPUT

Use SET STATISTICS IO ON to verify what actually happens
when the query executes
Slide 6 - 16
Optimization vs. Execution
Scans planned at optimization may not happen
update pt_tx set id = id + 1
Example 1 (scan planned and executed)

STATISTICS IO OUTPUT
Slide 6 - 17
Slide 2 of 5

STEP 1
|ROOT:EMIT Operator
|
| |UPDATE Operator
| |
| | |SCAN Operator
| | | FROM TABLE
| | | pt_tx
| | | Table Scan.
| | | Forward Scan.
| |
| | TO TABLE
| | pt_tx
STATISTICS IO OUTPUT
Slide 6 - 18
Table: pt_tx scan count 1, logical reads: (regular=6 apf=0 total=6), physical
reads: (regular=0 apf=0 total=0), apf IOs used=0
Table: pt_tx scan count 0, logical reads: (regular=0 apf=0 total=0), physical
reads: (regular=0 apf=0 total=0), apf IOs used=0
Total writes for this command: 9

update pt_tx
set id = id + 1
where 1 = 2
SHOWPLAN
Example 2 (scan planned but not executed)

In this example, the optimizer still plans a table scan
even though the WHERE clause short-circuits the
query at execution time
Use STATISTICS IO output to verify a scan really
happened
Slide 6 - 21
Continued
STATISTICS
Table: pt_tx scan count 1, logical reads: (regular=6 apf=0 total=6), physical reads: (regular=0 apf=0
total=0), apf IOs used=0
Table: pt_tx scan count 0, logical reads: (regular=0 apf=0 total=0), physical reads: (regular=0 apf=0
total=0), apf IOs used=0
Total writes for this command: 9

 Index types
 Optimizer selection criteria
 When indexes slow
access
 Index statistics and usage
Slide 6 - 22
Index Selection
Topics

ASE provides three types of indexes
Clustered
Nonclustered
Full text
One clustered index per table
Data is maintained in clustered index order
248 nonclustered indexes per table
Nonclustered indexes maintain pointers to rows
Full text indexing is beyond scope
Slide 6 - 23
Index Types

Indexes are balanced b-tree structures
An index may contain between 1 and 31 columns, but the
total index entry width must be no greater than
600 bytes
Text, image and bit columns can be indexed
Indexes are maintained and used internally by the server
to improve performance or to enforce uniqueness
The application programmer does not usually refer to
indexes
Slide 6 - 24
Index Types

Here, we’re clustered
on last name
With a clustered
index, there will be
one entry on the
last intermediate
index level page
for each data page
The data page is the
leaf or bottom
level of the index
Slide 6 - 25
Clustered Index Mechanism
Houston
Exeter
Brown
Albert
Quincy
Mason
Jones
Albert
Loon
Klein
Jude
Jones
Paul
Parker
Neenan
Mason
Alexis, Amy, ...
Root Page
Intermediate Page
Data Page
Amundsen, Fred, ...
Baker, Joe, ...
Best, Elizabeth, ...
Albert, John, ...
Masonelli, Irving, ...
Narin, Mabelle, ...
Naselle, Juan, ...
Neat, Juanita
Mason, Emma, ...
...
...
...

 Here, we have a
nonclustered
index on first
name
 The nonclustered
index has an
extra, leaf level
for
page / row
pointers
 Nonclustered
indexes do not
affect data
placement
Slide 6 - 26
Nonclustered Index Mechanism
Dave
Bob
Amy
Zelda
Elizabeth
Elizabeth
George
George
Amy
...
...
...
...
...
...
Sam
Sam
Alexis, Amy, ...
Root Page
Intermediate Page
Data Page
Amundsen, Fred, ...
Baker, Joe, ...
Best, Elizabeth, ...
Albert, John, ...
Masonelli, Irving, ...
Narin, Anabelle, ...
Naselle, Amy, ...
Neat, Juanita
Mason, Emma, ...
Zelda
...
...
...
Amy
Amy
...
...
Emma
...
Leaf Page
Anabelle
...

A clustered index tends to be 1 I/O faster than a
nonclustered index for a single-row lookup
Clustered indexes are excellent for retrieving ranges of
data
Clustered indexes are excellent for queries with
order by
Nonclustered indexes are a bit slower, take up much more
disk space, but are usually the next best alternative to a
table scan
Nonclustered indexes may cover the query for maximal
retrieval speed; For some queries; covered queries,
nonclustered indexes can be faster
Slide 6 - 27
Clustered vs. Nonclustered

A primary key is a logical concept, not a physical concept
Indexes are physical concepts, not logical concepts
There is a strong correlation between the logical concept of a key and the
physical concept of an index
By default, when you define relationships as part of table design, you will
build indexes to support the joins / lookups
By default, when you define a primary key, you will create a unique
clustered index on the table
Unique is good, clustered isn’t always good
Primary Key vs. Clustered
vs. Nonclustered
Page 6 - 28

During the index selection phase of optimization the
optimizer decides which (if any) indexes best
resolve the query
o Identify which indexes match the where clauses
o Estimate rows to be returned
o Estimate page reads
Slide 6 - 29
Optimizer Selection Criteria

 Indexes must correspond with SARGs.
 Useful indexes will specify a row or rows or set bounds for the
result set
 An index may be used if at least the first column of the index
matches the SARG
where dob between
'3/3/1941' and '4/4/65'
Slide 6 - 30
SARG Matching

Which of the following queries (if any) is helped by the index?
select * from authors
where au_fname = 'Jim' and au_lname = 'Smith‘
select * from authors where au_fname = 'Jim‘
select * from authors
where au_lname = 'Smith' or au_fname = 'Jim'
Slide 6 - 31
SARG Matching (Continued)
Create index a on authors(au_lname, au_fname)

Columns searched by range of values
Columns by which the data is frequently sorted (order
by or group by)
Sequentially accessed columns
Join columns (if other than the primary key)
Static columns
Slide 6 - 32
Using Indexes
Clustered Index Indications

NCI selection tends to be much more effective if less than
about 10% of the data is to be accessed (documentation
says 20%, it is wrong)
NCIs help sorts, joins, group by clauses, etc., if other
column(s) must be used for the CI
Index covering
Slide 6 - 33
Using Indexes
Nonclustered Index

The server can use the leaf level of a nonclustered index the way it
usually reads the data pages of a table: this is index covering.
The server can skip reading data pages
The server can walk leaf page pointers
A nonclustered index will be faster than a clustered index if the index
covers the query for a range of data (why?)
Adding columns to nonclustered indexes is a common method of
reducing query time
This has particular benefits with aggregates
Slide 6 - 34
Index Covering

Beware making the index too wide; As index width approaches row
width, the benefit of covering is reduced
# of levels in the index increases
Index scan time approaches table scan time
Remember that changes to data will cascade into indexes
Over-indexing can cause OLTP performance problems
Slide 6 - 35
Index Covering Continued

 What index will optimize this
query?
 What index will optimize this
query?
 In the second query, what
would the net effect be of
changing the range to this?
select title
from titles
where title = 'Future
Shock'
Slide 6 - 36
Index Selection Examples
select title
from titles
where price
between $5. and $20.
between $500 and $600

2,000,000 titles
40 rows / page (=50,000 pages)
1 million rows in the range
CI vs. NCI
Slide 6 -
select *
from titles
where price between $5. and $20.
Table facts:
INDEX USED PAGE READS
Clustered index
non-clustered index
25,000 +index levels
(worst case) 1,000,000,+ 7,150
No index (table scan) 50,000

It is feasible, occasionally likely, that a table scan is
faster than using a nonclustered index for specific
queries
The server evaluates all options at optimization time
and selects the least expensive query plan
Slide 6 - 38
CI vs. NCI

select title
from titles
or type = 'computing'
Slide 6 -
Or Indexing
 What indexes should (could) be used?
 Will a compound index help?
 Which column(s) should be indexed?

 How is the following query different (from a processing standpoint)?
 What is a useful index for?
Slide 6 -
Or Indexing Continued
select title
from titles
and type = 'computing'
select *
from authors
where au_fname in ('Fred', 'Sally')

 How many indexes may be useful?
SARG or SARG
Slide 6 - 41
Or Clauses
select *
from authors
where au_lname = 'Smith'
or au_fname = 'Fred'
select *
from authors
where au_lname in ('Smith', 'Jones', 'N/A')
Format
Examples

Table Scan
o Each row is analyzed, and criteria applied
o Matching rows are returned in the result set
o The cost of all the index accesses is greater than the
cost of a table scan
o At least one of the clauses names a column that is not
indexed, so the only way to resolve the clause is to
perform a table scan
Slide 6 -
Or Strategy
An OR clause may be resolved via a table scan, a multiple match
index or using OR Strategy

Multiple match index
o Using each part of the OR clause, select an index and
retrieve the row
o Only used if the results sets can not return
duplicate rows
o Rows are returned to the user as they are processed
Slide 6 -
OR Strategy Continued

OR Strategy (also known as dynamic index)
Used when the result from a multiple match index can
return duplicate rows
Any rowid that satisfies the OR clause is saved in a
temp table
The table is sorted
Duplicate rows are then eliminated
Results are then retrieved
Note: This strategy has additional overhead of repeated
reads and the creation and sorting of the work table
Slide 6 - 44
OR Strategy Continued

The server generates a dynamic index from indexes
supporting each individual SARG
The dynamic index consists of pointers to rows
matching one or both criteria
The server unions the results, removing duplicate
values
Note: If any individual SARG requires a table scan, a
table scan will resolve the query. Slide 6 - 45
OR: Example
select *
from authors
where au_lname in ('Baker', 'Garfield')
or state like 'A%'
Result

OR: Query Plan
select company, street2
from pt_sample
where id = 2017
or id = 2163
SHOWPLAN OUTPUT

 What is the best index?
 Do the columns being selected have a bearing on
the index?
Slide 6 - 48
Index Selection and the Select List
select *
from publishers
where pub_id = 'BB1111'

 Should there be a difference between the utilization
of the following two indexes?
Slide 6 -
Index Selection and the Select List
create index idx1 on titles (price)
/* or */
create index idx2 on titles (price, royalty)
select royalty
from titles
where price between $10 and $20

Composite (compound) indexes may be selected by
the server if at least the first column of the index is
specified in a where clause
Slide 6 -
Composite Indexes
create index idx1
on employee (division, department, emp_num)

Which queries will consider using the index?
Slide 6 -
Composite Indexes (Continued)
select *
from employee
where division = 'abc',
and department = 123
and emp_num = '123-456-789'
select *
from employee
where division = 'abc'
and emp_num = '123-456-789'
select *
from employee
where department = 123
and emp_num = '123-456-789'
create index idx1
on employee (division, department, emp_num)

Each additional index impacts update performance
In order to select appropriate indexes, we need to know
how many indexes the optimizer will use, and how many
rows are represented by the where clause
Slide 6 - 52
Composite vs. Many Indexes

CI or NCI on type
CI or NCI on price
One index on each of type & price
Composite on type, price
Composite on price, type
CI or NCI on type, price, pub_id, title, notes
 Which are the best options in which circumstances?
Slide 6 - 53
Composite vs. Many Indexes (Continued)
Options
select pub_id, title, notes
from titles
where type = 'Computer'
and price > $15.

It is imperative to be able to estimate rows returned for an index.
Therefore, the server will estimate rows returned before
index assignation
If statistics are available (When would they not be?) the server
estimates number of rows using distribution steps or index
density
If statistics are not available, the server estimates the number of
matching rows based on default statistics
o 10% for equality SARGS (=)
o 25% for closed range SARGS (between, > and <)
o 33% for open range SARGS (>, >=, <, <=)
If you have an equality join on a unique index, the server knows
only one row will match and doesn't need to use statistics
Slide 6 - 54
Index Usefulness

Improvised SARGs are WHERE clause restrictions that are not
SARGs at compile time, but are SARGs at run time
Adaptive Server will attempt to improvise SARGs
where it can
Adaptive Server will calculate indexes for improvised SARGs as
though they were literal SARGs
Adaptive Server will use table scans for those WHERE clause
restrictions it cannot improvise into SARGs
Prior to ASE 15 COLUMN = @VARIABLE is not a SARG at compile
time, so it cannot use distribution steps, it uses index density
Consider this example: what is the potential impact of an improvised
SARG?
Slide 6 - 55
WHERE with Improvised SARGs
COLUMN like @VARIABLE

These are SARGS at run time, but not at
optimization time. The server will treat this as a
SARG but will use index density or default
statistics, rather than distribution steps
Slide 6 - 56
When Distribution Steps Are Not Used
Constant Expression (prior to 15)
Local Variable as Constant
Col = 5*12 Col = @value

Selectivity is based on density stored in the statistics page
Selectivity = 1/# rows in smaller table if no statistics are available
o 1,000,000 orders, 5,000 customers
o selectivity = 1/5000
Probable number of orders/customer:
o 1,000,000 * 1/5000 = 200
Density = 1/number of unique values
o 1,000,000 rows, 1,000 unique values
o Density = 1/1,000
A single search value would, on average, yield 1,000,000 * .001
or 1000 rows
Low density is most selective; as density approaches 1, the index
approaches uselessness
Slide 6 -
Row Estimates

If there is no index, there will be a table scan, and the
estimate will be the number of pages in the table
(stored in sysindexes)
If there is a clustered index, estimate will be the number of
index levels plus the number of pages to scan
For a nonclustered index, estimate will be index levels +
number of leaf pages + number of qualifying rows
(which will correspond to the number of physical pages
to read)
For a unique index and an equality join, the estimate will
be 1 plus the number of index levels
Slide 6
Estimating Logical Page I/O

You have a 1,000,000 row table. The unique key has a
range (and random distribution) of 0 to 10,000,000
 How many rows will be returned by the following
query:
Slide 6 - 59
Data Distribution
select *
from table
where key between 1,000,000 and 2,000,000
How does the optimizer know whether to use an index or table scan?

 Adaptive Server keeps distribution information about indexes in
a table called sysstatistics
 There is distribution information for every index, and possibly
(depending on your DBA’s decision) statistics for nonindexed
columns as well
 The optimizer uses this information to estimate the number of
rows returned for a query
 The distribution page(s) are built at index creation time and not
maintained by the server
 The dbo should periodically issue the command to enable the
optimizer to continue picking the correct strategies
Slide 6 -
Index Statistics
update index statistics table [index]

With every release of the server, the optimizer gets
better at selecting optimal query paths
Forcing the optimizer to behave in a specific manner
does not allow it the freedom to change selection as
data skews
It also does not permit the optimizer to take advantage of
new strategies as advances are made in the server
software
Slide 6 - 61
When to Force Index Selection
Don't Do It!

When you (the developer) have information about a table
that Adaptive Server will not have at the time the query
is processed (i.e., using a temp table in a nested stored
procedure)
Occasions when you've proven the optimizer wrong
Slide 6 -
When to Force Index
Selection Continued
Exceptions

To force the server to use a specific index for a specific
table, you must first know the index id of the index
you want to use
In this example, the titles index with name pix will be
used for the titles table, and the publishers index
with name pix will be used for publishers
Slide 6 - 63
How to Force Index Selection
select *
from titles (index pix) join
publishers (index pix)
on titles.pub_id = publishers.pub_id

Partition strategies
Hash partitions
List
Range
Round robins
Using alter table
Add partition
64
Partition Support Topics

Divides large tables and indexes into manageable pieces
Are database objects that are managed independently
Create index cannot be done at partition level
Selects, inserts, deleting data can be done on partitioned table
Horizontal partitioning selects table rows that can be distributed among partitions on
different disk devices
Semantic partitioning allows data values in specified, key columns in each row to
determine the partition assignment of that row
Round-robin partitioning assigns rows randomly without reference to data values
65
Semantic Partition Support

Hash partitioning (semantic) –system supplied hash
function determines the partition assignment for each
row
List Partitioning (semantic) compares values in key
columns to user-supplied values for each partition
Range Partitioning (semantic) –compares key columns
within a set lower and upper bound for each partition
Round-robin partitioning – randomly assigned rows to
partitions so that each partition contains equal number
of rows (default)
66
Partition Strategies

Allows
Using the create table and create index commands when creating
partitions
Using alter table command to alter table’s partitioning strategy
Using add partition to partition to an existing table
Use of partitioning to expedite the loading of large amounts of data
67
Partition Strategies II

Round-Robin
Parallel Inserts
Parallel Query/DBCC
Create Index
Worker Threads
Segment
Partitions
Pre – ASE 15 Partitioning  Segment
Slices
Segment partitioning
o aka 'slices'
o Table only (no index
partitioning)
Primary goals
o Decrease last page
contention
o Allow parallel query
o Allow parallel dbcc
checkstorage
o Allow parallel index creation
Assessment
o Myth that they are no longer
useful due to SAN disk
speeds
o Datarow + partitioned table
is fastest for inserts (20-
25%)
o Not a great solution
But still decent

ASE 15.X Semantic Partitioning
Semantic data partitioning strategies
o Hash based partitioning
o Range Partitioning
o List Partitioning
o Round-Robin Partitioning
Index partitioning
o Global indexes – Index spans all partitions
o Local indexes – Index spans one partition
Improved query support
o Optimizer and execution support
(parallelism & elimination)
Partition-aware maintenance
o Update statistics on one or all partitions
o Truncate, reorg, dbcc, bcp (out) partition
Range/Hash
Parallel Inserts
Parallel Query/DBCC
Create Index
Worker Threads
Range/Hash
Partitions
A-I J-R S-Z

Range Partitioning - Example
create table customer
( c_custkey integer not null,
c_name varchar(20) not null,
c_address varchar(40) not null,
other columns …
) partition by range (c_custkey)
(cust_ptn1 values <= (20000) on segment1,
cust_ptn2 values <= (40000) on segment2,
cust_ptn3 values <= (60000) on segment3 )
cust_ptn3:
values <=60000
Segment 3
cust_ptn2:
values <=40000
Segment 2
cust_ptn1:
values <=20000
Segment 1

Indexes on a partitioned table may be partitioned
o Hash, List, Range only
o Round-Robin does not support partitioned indexes
Think about it
o If partitioned, an index inherits the same partition keys as the table
Terminology
o Local Index  Index is partitioned
o Global Index  Index is not partitioned (single b-tree)
PKey Constraint is automatically partitioned
o This can lead to pkey enforcement issues if the partitioning is on columns
other than pkey
You will get a warning about this if it applies
o Work-around is to use a unique non-clustered (global) index on PKey vs.
Pkey constraint.
Index Partitioning

Clustered (Local) Index -
ExampleCustomer is range-partitioned table on c_custkey column
Create unique clustered index ci_nkey_ckey on
customer(c_custkey, c_nationkey)
Segment 2 Segment 3Segment 1
Note: All primary key constraint indexes for range, hash, list partitioned
tables are local indexes by definition.

(Non-clustered) Global Index
Create unique index ci_nkey_ckey
on customer(c_nationkey,
c_custkey)
on segment4
Segment 3Segment 2Segment 1
Segment4

(Non-clustered) Local Index
Create unique index ci_nkey_ckey
on customer(c_nationkey, c_custkey)
on segment4 local index
Segment 1 Segment 2 Segment 3
Segment4

Performance
Enhancement Through
Partitioned IndexIncreased concurrency through multiple index access points.
o Reduced root page contention
Index size adjusted according to the number of rows in each partition
Fewer index pages searched for smaller partitions
15.x Local Index on partitioned tablePre-15.0 with unpartitioned index
Query A
Query B
Query C
Unpartitioned Table Partitioned Table
Query A Query B Query C
Partitioned Index
Shorter Access Path
Unpartitioned Index

Uniqueness may not be enforceable!!!
o If you partition a table by keys other than the pkey columns, duplicate
key values may be inserted
o ASE will warn you of this when you partition the table
Local Indexes vs. Unique Indexes
CACA NYNY TXTX
Partitioned Table by State (pkey is cust_id)
Insert table (cust_id, state, …)Insert table (cust_id, state, …)
values (12345,'CA', …)values (12345,'CA', …)
values (12345,'NY', …)values (12345,'NY', …)
values (12345,'TX', …)values (12345,'TX', …)
Local Index (Partitioned by State)
Access path

Why does ASE Just warn vs. enforce?
o Originally table is unpartitioned with indexes 7 levels deep
o Now, partition on state
Assume 10 partitions
Assume now each index is only 3 levels deep
Unique enforcement now requires 30 (10x3) I/O's vs. 7
Unique enforcement would block partition specific maintenance
What's the Work-Around?
o Create unique/pkey indexes as nonclustered global indexes
o Other indexes can still be local – helping insert speeds.
Local Indexes vs. Unique
Indexes

Function-based indexes permit indexes based upon derived values in
multiple columns
Function-based indexes
Slide 6 -
create index nci on invoice_header
(substring (ID,2,9))

The optimizer uses indexes to improve query performance when possible
Watch out for improvised SARGs
Queries with OR may require a table scan
Try to take advantage of covered queries
Be careful when forcing an index
Partition allows large tables to be broken into usable pieces
Partition Strategies
Hash partitioning determines the partition assignment for each row
List Partitioning supplied values for each partition
Range Partitioning compares key columns within a set lower and upper bound
for each partition
Round-robin partitioning – randomly assigned rows to partitions so that each
partition contains equal number of rows (default)
Slide 6 - 79
Summary

1. Use showplan to observe what ranges of values tend to use an index
versus a table scan to resolve this query
2. Now declare variables @low and @high and see whether variations in the
values are reflected in the optimization
3. Does it make any difference if you do this in a stored procedure?
4. Use set statistics io to track reads with various ranges, then force use of
the index / scan technique and track the query statistics. Can you
outguess the optimizer?
Slide 6 -
Lab 6.1: Indexes vs Table Scans
select max(id)
from pt_tx_NCamount
where amount between ? and ?

Annual Conference, 2015 (c) 2015 Independent SAP Technical User Group
QuestionsandAnswersQuestionsandAnswers

Annual Conference, 2015 (c) 2015 Independent SAP Technical User Group
Thank You for AttendingThank You for Attending
Pleasecompleteyour session feedbackPleasecompleteyour session feedback
formform

Choosing Indexes For Performance

More Related Content

What's hot (20)

Similar to Choosing Indexes For Performance (20)

More from SAP Technology (20)

Choosing Indexes For Performance