T HE F ORGOTTEN F ILL FACTOR Optimize Performance Minimize Maintenance Reduce Problems.

THE FORGOTTEN FILL FACTOROptimize Performance

Minimize Maintenance

Reduce Problems

ABOUT ME

SQL Server MCT Member of the Belgian Microsoft Extended

Expert Team

PURPOSE OF THIS SESSION

WHAT IS A FILLFACTOR

Optional INT value

Specifies empty pages in an Index

Server-wide default is 0

VISUAL

Good Fill Factor Bad Fill Factor

IMPACT ON PERFORMANCE

ADVISE OR HELP?

The result of this is: Try and Error “Forget” it “Default” it “Maintain” it No/Bad Indexes

Some exception examples: Very good whitepaper by Ken Lassesen Good blog post by Pinal Dave

WHY SHOULD WE CARE?

Increased Read Times Increased Write Times Increased CPU usage Index becomes useless Full Table Scans Index Maintenance comes under pressure Bad balance between main resources ! Compression Increases the problem !

NOW LETS START THE DIVE TO LEVEL 400

LET’S FOCUS ON READ TIME

Serial Read times: 0,5 ms Random Read times: Between 3,5 and 15 ms 1 non serial read, delays read speed with

13% 98 * 0,5 + 2*3,5 = 56 ms vs. 50 ms.

(98 serial reads, move to the disjoint one + move back)

• Less important with SSD’s!

READS - LET’S USE SOME MATH I

Fragments

Adjacent Read Time

Disdjoint read time

Increase

0 50 0 0 %

1 49,5 7 13 %

2 49 14 26 %

3 48,5 21 39 %

4 48 28 52 %

5 47,5 35 65 %

10 45 70 130 %

15 42,5 105 195 %

20 40 140 260 %

40 30 280 520 %

Fill Factor

# Pages

Total Read Time

Increase

100 100 50 0 %

99 102 50,5 2 %

98 103 51 3 %

97 104 51,5 4 %

96 105 52 5 %

95 106 52,5 6 %

90 112 55 12 %

85 118 57,5 18 %

80 125 60 24 %

60 150 70 48 %

Cost of Fragmentation

(100 Pages)

Cost of higher Fill Factor

(100 Pages)

LET’S FOCUS ON WRITE TIMES

Similar behavior – Depends on Buffer Time to write to one extent : 0,3 ms Sequential write is 70% faster then Random

in Throughput Impact is even bigger on SSD’s The larger the subset, the bigger the impact

REAL LIFE EXAMPLES

No tuning (Only) Fill Factor Tuning(Only) Fill Factor Tuning

LET’S START THE CALCULATIONS

PREREQUISITES

We will focus on Read times Focus on OLTP Databases All other parameters will benefit as well

But increase is more difficult to calculate Is to dependent from incalculable parameters

Focus Highest possible Fill factor Lowest possible Fragmentation Highest possible Page fill ratio Acceptable Maintenance

NARROW DOWN THE PROBLEM INDEXES

Only tables that are large enough (> 8MB) Skip all Indexes where first Key is

Monotonically increasing Focus on:

Default Fill Factor High Fragmentation Average Fragmentation, High Page fill Low Fragmentation, Low Page fill Fill Factor < 100, Read Only Partition

IDENTIFY YOUR INDEX KEYS PER TYPE

THE EASY ONES…SINGLE KEY INDEXES - MONOTONIC

Monotonic Increasing Keys Identity Timestamp Rowversion

Careful with Date

Only if not assigned by code, but assigned by function

Rows containing extendable data types If these fields are updated and become larger, a

page split will occur

Fill Factor should always be 100% Empty pages will never be used

THE OTHERS…WILL HAVE FRAGMENTATION

And will need maintenance!

FIRST ACTIONS (READ AS QUICK-FIX)

Use Table Partitioning where possible Enterprise & Not every partition needs the same fill factor

single_partition_rebuild_index_option Offline!

Use filegroup/Database growth as initial guess Backup’s will give you an initial figure

You will need exact figures later on though

Narrow down to the problem Indexes Write a sys. Query to find:

Schema Name; Table Name; Index Name Key size, Index Size Fragmentation, Page Space Usage Current Fill Factor ! Partition !

CALCULATE POSSIBLE FILL FACTORS

What’s needed Key Size

How FLOOR(8060 / Key Size) = # Possible values

Calculates the array with possible FF Values Example

Int key (4 bytes) => 2015 possible FF values Only 100 are available for usage

KEY INDEXES – (SEMI)-SEQUENTIAL

Because they sometimes behave predictable We can still use statistics

Calculate possible fill factors Estimate the fragmentation likeliness Plot the possible fill factors vs. the Fragmentation

likeliness Achieve low fragmentation without wasting to much

space.

KEY INDEXES – (SEMI)-SEQUENTIAL

Sequential, but non unique Semi-Sequential, but unique Semi-Sequential, Non unique Key Features

High Selectivity. Gets inserted out of order with small derivations.

Or has multiple entries for the same key. Behaves predictable (Gaussian)

CALCULATE MAX POSSIBILITY OF FRAGMENTATION

Non unique sequential Sample the Key values

Select Count(Key),Key From Table Group by Key Order by Count(Key) Desc

Returns the maximum possible occurrence (Collisons)

Unique Semi-sequential Do we have a sequential Key? Use it to find the max out of sequence key values

(Read Fragmentation) Else

Do we have the out of sequence probability ratio? Can be used as initial growth ratio

Treat it as random

SIMPEL SEQUENTIAL CALCULATIONS

(8060)/([KeySize]) = #Entries/page Be carefull with

nullable datatypes Expandable Datatypes

[KeySize]*[MaxOccurence] = MaxEntriesPerKey

([KeySize]*[MaxOccurence])*(1-MaxFragmentation)=BestEntriesKey

INDEXES – RANDOM / CHAOS

Now it gets interesting We need Index data, to tune the index itself Key Indicators (no points for guessing)

The current fill factor Capability to cope with randomness & growth

Current Index Page fill ratio Current page usage Effectiveness of growth prediction

Current index Fragmentation level Indication of the real randomness & Growth

Current Table Growth ratio / Maintenance interval Amount of Rows as this influences the growth ratio Data type of the first column of the Index

Example GUID vs. Int

CALCULATE MAX SUPPORTED GROWTH RATIO FOR A SPECIFIC FILL FACTOR

DEMO TIMELET’S GO INTERACTIVE

NEVER FORGET

Page fill ratio Indication of effectiveness

Fill Factor improves Insert speeds But

Is badly used if the forseen space isn’t used! If badly used will

Increase read times Increase storage usage Decrease performance

Optimise for insert Be carefull with rebuilds

Partition Optimise for read

TIME FOR DECOMPRESSION