James SnapeApplication Development ConsultingMicrosoft Limited

Microsoft Analysis Services Physical Design

The Kimball Process

Agenda

Hardware

Dimensions

Facts

Relational stuff

Performance tuning next steps

NB: Relational design not complete –logging, auditing etc discussed in next session

Prime Directive:

Sequential IO Good,

Random IO Bad

Hardware

SQL Server Fast Track Data Warehousewww.microsoft.com/sqlserver/2008/en/us/fasttrack.aspx

Pre-tested hardware configurations

Specific disk, filegroup, layouts

Minimal indexing

To feed CPU at maximum capacity

Dimensions vs Facts

Dimension

Small (relatively)

Repeating data

Fact

Large

Numeric data + keys

Treat them differently

Dimensions in Relational Terms

Table structure

Keys

Indexes

Null handling

Managing change

Processing

CustomerFull Name

Post Code

City

State

Country

Gender

Occupation

Marital Status

Email Address

Customer

Geography1. Country

2. State

3. City

4. Post Code

5. Full Name

Star vs. Snowflake Schemas

dbo.CustomerCustomerKey

FullName

PostCode

City

State

Country

Gender

Occupation

MaritalStatus

EmailAddress

dbo.CustomerCustomerKey

GeographyKey

FullName

Gender

Occupation

MaritalStatus

EmailAddress

dbo.GeographyGeographyKey

PostCode

City

State

Country

OR

NB: both are denormalized, one more than the other

Primary Keys

Use smallest possible integer as surrogate primary key

Primary key is a “row identifier”

Multiple row “versions” are possible

“None” and “Unknown” special values are useful

Do NOT use business/source system keys

Clustered primary key is OK for dimensions

Dimension Indexes

Dimension processing queries of the form:

SELECT DISTINCT .... FROM ....

WHERE (filter) clauses never used

WHERE (join) clauses are used in snowflake dimensions

Non-processing queries may end up in SQL

ROLAP dimensions

Direct to SQL queries

Null Handling in Dimensions

By default NULL converts to 0 or an empty string

NULL attribute keys can invoke special “Unknown Member” handling

Prefer to create a specific “Unknown” row

CustomerKey FullName City Country

-1 Unknown Unknown Unknown

-2 None None None

1243 John Smith London United Kingdom

1244 Mary Jones Glasgow United Kingdom

Dimension Attributes

Attributes have keys, names (and values)

Integer attribute keys are smaller and faster

Keys must be unique

SELECT [Month] as [Month],

[Month] + „ „ + [Year] as [Month of Year]

FROM dbo.Time

Attribute Key Name (Value)

Year 2009 CY 2009 2009

Month 4 April 4

Month of Year 20090400 April 2009 4

Slowly Changing Dimensions

PK = row identifier

Multiple rows = multiple versions

Add effective dating columns

Which can be exposed as new dimensional attributes

dbo.CustomerCustomerKey

FullName

PostCode

City

State

Country

Gender

Occupation

MaritalStatus

EmailAddress

EffectiveFrom (smalldatetime)

EffectiveTo (smalldatetime)

CurrentFlag (tinyint)

Facts in Relational Terms

Keys

Indexing

Partitioning

Processing

Consider Row and Page compression

Internet SalesSales Amount

Order Quantity

Tax Amount

Unit Price

Transaction Count

Fact Keys and Indexes

Is a surrogate/primary key required?

Beware the clustered index/primary key

Prefer the date FK as the clustered index

Add NO CHECK to foreign keys

Indexes are usually not useful

Unless processing degenerate dimensions

Or servicing ROLAP/direct to SQL queries

Fact Partitioning – Why?

Parallel processing

Only process most recent data

Multiple storage engine threads during query

Archive off data

Multiple aggregation strategies

NB: Partitions require Enterprise Edition

Fact Partitioning – Guidelines

Partition when fact tables are 50-100GB+

Ideal partition size 2M-20M rows

Less than 1000 partitions per measure group

This wins over partition size

Prefer to partition over time

Can not aggregate higher than partition grain

Align AS and SQL partitions!

Calculated time keys become very useful

Fact Storage

MOLAP, ROLAP or HOLAP

Source Data Facts Aggregations

Relational Multidimensional

Proactive Caching

Cube = “Cache”

Automatic invalidation of cube

Automatic rebuild of cube

Valid?Valid?

Query

SQL Query

Quick Storage Engine Tuning

Ensure attribute relations are implemented

Turn on query log

Run Usage Based Optimisation (UBO) wizard

© 2007 Microsoft Corporation. All rights reserved. Microsoft, Windows, Windows Vista and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other

countries.

The information herein is for informational purposes only and represents the current view of Microsoft Corporation as of the date of this presentation. Because Microsoft must respond to

changing market conditions, it should not be interpreted to be a commitment on the part of Microsoft, and Microsoft cannot guarantee the accuracy of any information provided after the date of

this presentation.

MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.