IBM® DB2® for Linux®, UNIX®, and Windows®
®
Best practices
Optimizing analytic workloads using
DB2 10.5 with BLU Acceleration
Jessica Rockwood
Senior Manager, DB2 LUW
Performance Benchmarking,
IBM Canada Lab
Roman B. Melnyk
Senior Information Developer, IBM
Canada Lab
Michael Kwok
Senior Manager, DB2 LUW
Warehouse Performance, IBM
Canada Lab
Berni Schiefer
Distinguished Engineer, Information
Management Performance and
Benchmarks, IBM Canada Lab
Issued: September 2013
Executive summary ............................................................................................. 4
Introduction .......................................................................................................... 5
Before you start .................................................................................................... 6
Hardware and software ................................................................................ 6
Identifying optimal workloads for BLU Acceleration ................................................7
Capacity planning................................................................................................ 7
Processor cores ............................................................................................... 7
Memory ........................................................................................................... 8
Input/Output .................................................................................................. 8
Best practices for configuration ......................................................................... 8
Single setting for analytic workloads.......................................................... 8
Database and database manager configuration ...................................... 10
Adopting BLU Acceleration ............................................................................. 11
Guidelines for creating column-organized tables ................................... 12
Synopsis tables and data skipping ............................................................ 13
Loading data ................................................................................................. 14
Query execution plans and the new CTQ operator ................................ 17
Database maintenance................................................................................. 19
Concurrency control (workload management) ....................................... 19
Monitoring .................................................................................................... 20
Is the sort heap adequate for good performance? .....................................................20
Is the table suitably organized?....................................................................................21
How well are queries performing?..............................................................................21
How is the buffer pool performing?............................................................................22
How are the prefetchers performing? .........................................................................22
Conclusion .......................................................................................................... 23
Further reading................................................................................................... 24
Contributors.................................................................................................. 24
Acknowledgements ..................................................................................... 24
Notices ................................................................................................................. 26
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 2 of 27
Trademarks ................................................................................................... 27
Contacting IBM .................................................................................................. 27
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 3 of 27
Executive summary
DB2 10.5 with BLU Acceleration is a combination of complementary innovations from
IBM that simplifies and speeds up analytic workloads. It is easy to implement and is selfoptimizing. BLU Acceleration eliminates the need for indexes, aggregates (for example,
MQTs or materialized views), or time-consuming database tuning to achieve top
performance and storage efficiency. No SQL or schema changes are required to take
advantage of this breakthrough technology.
These innovations, which are introduced in DB2® for Linux®, UNIX ®, and Windows®
Version 10.5 (DB2 10.5), are designed to help you quickly find answers to more complex
business questions while keeping costs down.
DB2 10.5 with BLU Acceleration offers a rich set of features that help you to meet these
goals, including column-organized storage, actionable compression, parallel vector
processing, and data skipping. In combination, these features provide an in-memory,
CPU-optimized, and I/O-optimized solution.
This paper gives you an overview of these technologies, recommendations on hardware
and software selection, guidelines for identifying the optimal workloads for BLU
Acceleration, and information about capacity planning, memory, and I/O. A section on
system configuration shows you how IBM’s focus on simplicity enables you to set up
DB2 10.5 so that it automatically makes optimal configuration choices for analytic
workloads. Other sections describe how to implement and use DB2 with BLU
Acceleration. The focus of this best practices paper is on ease of use. You will learn how
BLU Acceleration works and what it is doing “under the covers”, which will give you a
real appreciation of the simplicity that is built into BLU Acceleration and show you that it
really does deliver “super analytics, super easy”.
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 4 of 27
Introduction
BLU Acceleration is a new collection of technologies for analytic queries that are
introduced in DB2 for Linux, UNIX, and Windows Version 10.5 (DB2 10.5 1 ). At its heart,
BLU Acceleration is about providing faster answers to more questions and analyzing
more data at a lower cost. DB2 with BLU Acceleration is about providing order-ofmagnitude benefits in performance, storage savings, and time to value.
These goals are accomplished by using multiple complementary technologies, including:

The data is in a column store, meaning that I/O is performed only on those
columns and values that satisfy a particular query.

The column data is compressed with actionable compression, which preserves
order so that the data can be used without decompression, resulting in huge
storage and CPU savings and a significantly higher density of useful data held in
memory.

Parallel vector processing, with multi-core parallelism and single instruction,
multiple data (SIMD) parallelism, provides improved performance and better
utilization of available CPU resources.

Data skipping avoids the unnecessary processing of irrelevant data, thereby
further reducing the I/O that is required to complete a query.
These and other technologies combine to provide an in-memory, CPU-optimized, and
I/O-optimized solution that is greater than the sum of its parts.
BLU Acceleration is fully integrated into DB2 10.5, so that much of how you leverage
DB2 in your analytics environment today still applies when you adopt BLU Acceleration.
The simplicity of BLU Acceleration changes how you implement and manage a BLUaccelerated environment. Gone are the days of having to define secondary indexes or
aggregates, or having to make SQL or schema changes to achieve adequate performance.
This best practices paper focuses as much on “what you no longer need to do” as on
“what you should do”. We direct your attention to a few key implementation items, and
then you can let BLU Acceleration take over and provide you with optimal performance.
It is analytics that is super fast and super easy – just load and go!
1
This document is current with respect to DB2 10.5 FP1 (DB2 10.5.0.1) and later. The use of the most current fix pack is
encouraged.
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 5 of 27
Before you start
Ensure that your system meets the necessary hardware and software requirements.
Hardware and software
BLU Acceleration is supported on POWER®/AIX® and x86/Linux platforms. Table 1
summarizes the supported operating systems, including both minimum and
recommended versions, as well as recommended processors. Note that the minimum
version requirements on these platforms are the DB2 10.5 requirements, which are
applicable to both row- and column-organized tables.
Table 1. Hardware and operating system recommendations
Operating
system
Minimum version
requirements
Recommended
versions
Hardware
recommendations
AIX
AIX 6.1 TL7 SP6 or
AIX 7.1 TL1 SP6
AIX 7.1 TL2 SP1 or
higher
POWER7 or higher 2
Linux x86
(64-bit
only)
Red Hat Enterprise
Linux (RHEL) 6, SuSE
Linux Enterprise
Server (SLES) 10 SP4
or SLES 11 SP2
RHEL 6.3 or higher
Intel Nehalem (or
equivalent) or higher 3
SLES11 SP2 or higher
For optimal performance, we recommend that you maintain your system firmware at the
latest levels, particularly if you are using virtualization.
BLU Acceleration is offered in DB2 Advanced Workgroup Server Edition, DB2
Advanced Enterprise Server Edition, and DB2 Developer Edition.
2
3
Any POWER processor that supports DB2 10.5 is supported, but there are specific hardware optimizations on this processor.
Any Intel processor that supports Linux x86 is supported, but there are specific hardware optimizations on this processor.
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 6 of 27
Identifying optimal workloads for BLU Acceleration
In general, column-organized tables significantly speed up analytic workloads or data
mart types of workloads. A data mart or warehouse typically has queries that have
grouping, aggregation, range scans, or joins; queries that access a subset of a table’s
columns; and database designs that often include a star or snowflake schema.
DB2 10.5 supports using both row-organized and column-organized tables in the same
database, even in the same table spaces and buffer pools. The benefit of having column
organization and row organization in the same database is that you can choose to
optimize the table layout based on the workload. Row-organized tables have long been
optimized for online transactional processing (OLTP) environments. Now with columnorganized tables sitting right next to row-organized tables, you have the best of both
worlds in the same database.
In fact, IBM provides Optim™ Query Workload Tuner 4 to identify and validate tables
that would benefit from column organization based on the application workload.
Capacity planning
This section provides some guidelines for the number of cores and the amount of
database memory that are recommended to achieve optimal performance for a BLUaccelerated workload.
Processor cores
DB2 with BLU Acceleration exploits multi-core parallelism and leverages SIMD
parallelism to accelerate the performance of analytic queries. The effectiveness of these
optimizations is directly impacted by the number of cores that are available to DB2 with
BLU Acceleration.
A minimum of 8 processor cores is required 5 for BLU-accelerated workloads. Today a
single processor chip typically has 8-12 cores. If the DFT_DEGREE database
configuration parameter has the default setting of ANY, queries use a degree of
intrapartition parallelism that is equal to the number of cores. When several queries are
running concurrently, the degree is automatically scaled back, based on the number of
active agent threads on the system, to maintain efficient execution.
Higher data volumes, higher degrees of query concurrency, and greater query
complexity would benefit from a larger number of cores.
4
Optim Query Workload Tuner is part of DB2 10.5 AESE and AWSE
http://pic.dhe.ibm.com/infocenter/prodguid/v1r0/clarityreports/report/html/softwareReqsForProductByComponent?deliverableId=254B4BA0C5F011E18183F12B0925FE36&duComponent
=Server_8388E5E04AD611E2A6D822020925FE1B
5
Issued: September 2013
Memory
DB2 with BLU Acceleration is designed to effectively leverage large memory
configurations. To realize the greatest benefit from the in-memory analytics that is built
into DB2 10.5, consider systems that have at least 64 GB of RAM for production use. As
you scale up the system maintain a ratio of 8GB of RAM per core.
For information on how to allocate this memory within DB2, see "Best practices for
configuration".
Input/Output
BLU Acceleration has more random I/O access than in row-organized data marts that
typically perform large sequential scans. Consider storing key column-organized tables
on I/O subsystems and devices that provide high random I/O access. Examples of storage
that provides better random I/O access are solid state drives (SSD),flash-based storage or
enterprise SAN storage systems BLU-accelerated tables that are frequently accessed
benefit from improved I/O subsystem performance, particularly when the table exceeds
the memory size that is available for use by BLU Acceleration.
There are both internal and external SSD or flash solutions. Internal flash storage might
not have any write cache, and that can affect write operations involving, for example,
temporary tables that are created and populated during query processing.
Common examples of storage that would provide good random I/O characteristics for
use with DB2 with BLU Acceleration include FlashSystem (810/820), Storewize V7000
(with SSD), XIV, as well as enterprise SAN storage such as DS8K.
Best practices for configuration
The simplicity of managing a BLU-accelerated environment is one of the key benefits of
this technology. There is even a new option for the DB2_WORKLOAD registry variable
to automatically set configuration parameters that are most relevant to BLU Acceleration
and the performance of analytic workloads.
Single setting for analytic workloads
If you have a database that is dedicated to analytic workloads, you can use a simple
setting to configure DB2 and the database for BLU Acceleration.
By setting the DB2_WORKLOAD registry variable to ANALYTICS before you create the
database, DB2 is aware that the system will be used for analytic workloads. The new
database is automatically configured for optimal analytics performance. Remember to
not disable AUTOCONFIGURE in the CREATE DATABASE command. For example:
db2set DB2_WORKLOAD=ANALYTICS
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 8 of 27
db2start
db2 create db mydb
Be sure to set this registry variable (when applicable) before creating your analytics
database. DB2 will do the rest.
Table 2 describes in detail what happens to a newly created database when
DB2_WORKLOAD=ANALYTICS.
Table 2. Database behavior that is influenced by DB2_WORKLOAD=ANALYTICS
Category
Automatic best practice settings
General database configuration
parameters
DFT_TABLE_ORG = COLUMN
PAGESIZE = 32 KB
DFT_EXTENT_SZ = 4
DFT_DEGREE = ANY
Database memory configuration
parameters
CATALOGCACHE_SZ, SORTHEAP, and
SHEAPTHRES_SHR are set to a value that is
higher than the default and optimized for the
hardware that you are using.
Intraquery parallelism
Enabled for any workload (including
SYSDEFAULTUSERWORKLOAD) that specifies
MAXIMUM DEGREE DEFAULT, even if the
database manager configuration parameter
INTRA_PARALLEL = OFF.
DB2 workload management
The default concurrency threshold on the
SYSDEFAULTMANAGEDSUBCLASS service
subclass is enabled to ensure maximum
efficiency and utilization of the server.
Automatic space reclamation
Sets AUTO_MAINT = ON and AUTO_REORG =
ON, and space reclamation is performed for
column-organized tables by default.
The DB2_WORKLOAD registry variable is set at the instance level. It influences all new
databases in that instance. If the instance is going to have multiple databases, not all of
which are BLU-accelerated data marts, simply set the registry variable before you create
the databases that you want optimized for analytic workloads, then unset the registry
variable after these databases are created. Alternatively, follow the recommendations
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 9 of 27
below for optimizing a pre-existing database, and apply only the database-level
recommendations.
To optimize a pre-existing database by configuring the recommended settings, set the
registry variable and then issue the AUTOCONFIGURE command. The following
example shows you how to optimize the pre-existing database MYDB for analytics and
BLU Acceleration:
db2set DB2_WORKLOAD=ANALYTICS
db2start
db2 connect to mydb
db2 autoconfigure apply db
After auto-configuration completes (either as part of the CREATE DATABASE command
or an AUTOCONFIGURE command), review the sizes of the buffer pools, sort heap, sort
heap threshold, and utility heap. You can fine tune the memory sizes based on your
specific application needs and your expected data volumes. The following section
provides more details about this type of review.
Database and database manager configuration
BLU Acceleration uses database shared sorts (the memory size for which is configured by
the SHEAPTHRES_SHR database configuration parameter). In DB2 10.5 with BLU
Acceleration, the self-tuning memory manager (STMM) does not adjust the values of the
SHEAPTHRES_SHR and SORTHEAP database configuration parameters automatically.
Instead, these parameters are set with static values by the AUTOCONFIGURE command
when DB2_WORKLOAD=ANALYTICS.
At the instance level, the SHEAPTHRES configuration parameter is a soft target value
that DB2 tries to maintain for all memory that is consumed by both private and shared
memory-intensive operations, such as hash join and hash aggregation. This value is not
just about sort memory, but about all “working memory”. Keep the instance-level
SHEAPTHRES configuration parameter at its default value of 0, which leaves the
tracking of sort memory consumption at the database level only (SHEAPTHRES_SHR).
However, at the database level, the value of the SHEAPTHRES_SHR parameter
determines the limit for the allocation of sort memory. When a certain percentage of the
value of SHEAPTHRES_SHR is attained, the DB2 database manager begins to throttle the
memory that is allocated to sort memory consumers to prevent memory overcommitment, and those consumers might therefore receive only their minimum
allocation requirement.
To ensure that aggregation-intensive operations, which are common in analytic
workloads, have sufficient sort memory to perform optimally, take the following steps:

Set the SHEAPTHRES_SHR parameter to 40-50% of the total memory that is
available to the database, which is determined by the value of the
DATABASE_MEMORY database configuration parameter.
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 10 of 27

Set the SORTHEAP parameter to 5-20% of the value of the SHEAPTHRES_SHR
parameter.
With increasing query concurrency and complexity, decrease the SORTHEAP to
SHEAPTHRES_SHR ratio. In general, you can consider increasing the value of the
SHEAPTHRS_SHR parameter, but beware of substantially increasing the value of the
SORTHEAP parameter without understanding overall memory requirements across the
entire workload.
Table 3 summarizes what the memory distribution within the database should be from a
memory configuration standpoint. Your specific application and workload characteristics
might suggest different values.
Table 3. Recommended memory distribution for a BLU-accelerated database as a
function of workload concurrency
Low concurrency (< 20 concurrent
workloads)
High concurrency (> 20 concurrent
workloads)
40% buffer pool
25% buffer pool
40% SHEAPTHRES_SHR
50% SHEAPTHRES_SHR
SORTHEAP = SHEAPTHRES_SHR / 5
SORTHEAP = SHEAPTHRES_SHR / 20
Because of their complexity, analytic queries are more likely to benefit from an increased
value of the STMTHEAP database configuration parameter, which specifies the size of
the statement heap (a work space for the SQL compiler during compilation of an SQL
statement). If your query returns SQL0437W with reason code 1, consider increasing the
STMTHEAP value. Start by increasing the value of STMTHEAP by 50%; if the warning
persists, continue increasing the value by the same percent increment, as needed.
DB2 10.5 sets the value of the UTIL_HEAP_SZ database configuration parameter to
AUTOMATIC for new databases. Consider setting the UTIL_HEAP_SZ parameter to
AUTOMATIC for existing databases. This setting enables the automatic resizing of the
utility heap by DB2 load and other utilities.
Adopting BLU Acceleration
This section is all about implementing and using BLU Acceleration. It provides
recommendations on how to build and load column-organized tables, details about
column-organized tables and query execution, and guidelines for monitoring your
database.
If you want to convert one or more row-organized tables in your database into columnorganized tables, you can simply use the db2convert command to automate the
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 11 of 27
process. This command calls the ADMIN_MOVE_TABLE stored procedure to perform
the conversion, and shares its options. The db2convert command drops any dependent
objects (such as secondary indexes or aggregates) as part of the conversion operation, as
shown in the following example, because they are not required on column-organized
tables:
$ db2convert -d mydb -z db2inst1 -t mytable
Conversion notes for exceptional table(s)
----------------------------------------Table: DB2INST1.MYTABLE:
-Secondary indexes will be dropped from the table.
Enter 1 to proceed with the conversion.
Enter 2 to quit.
Guidelines for creating column-organized tables
Creating a column-organized table is very easy. All you need is a table name, column
names, and their data types. If you set the DFT_TABLE_ORG database configuration
parameter to COLUMN, you do not need to specify the ORGANIZE BY COLUMN
clause. Here is an example of the syntax to define a column-organized table:
CREATE TABLE mytable (
c1 INTEGER NOT NULL,
c2 INTEGER,
…
)
ORGANIZE BY COLUMN IN <table_space_name>
No additional specifications are required. The previous example includes the IN
table_space_name clause. For improved manageability, it is recommended that you create
fact tables in their own table space. For example, with a fact table in one table space and
dimension tables in a different table space, you can have different storage groups and
buffer pools for each table space (and its associated table type) as well as independent
backup/restore fact tables. If there are enforced primary key constraints or unique
constraints, the unique indexes that enforce those constraints should also be in their own
table space, which you can specify by using the INDEX IN clause.
When possible, use NOT NULL as the default attribute for columns, except in cases
where you actually need to store null values. Every nullable column in a columnorganized table is actually a two-column group of columns: one for the nullability
attribute and one for the actual column. These nullable columns cause extra work to be
done on each row during query processing.
It is also useful to add primary key, informational foreign key, and any informational
check constraints after your initial data has been loaded.
To determine whether a table is defined as row-organized or column-organized, you can
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 12 of 27
query the SYSCAT.TABLES catalog view. A new column, TABLEORG, contains "C" for
column-organized and "R" for row-organized. The following example shows you how to
query the catalog view:
SELECT tabname, tableorg, compression
FROM
syscat.tables
WHERE tabname like 'SALES%';
TABNAME
------------------------------SALES_COL
SALES_ROW
TABLEORG COMPRESSION
-------- ----------C
R
N
2 record(s) selected.
The COMPRESSION column for column-organized tables is blank because the data in
such tables is always compressed.
You can define a primary key by using either the CREATE TABLE statement or an
ALTER TABLE statement if the table exists, just as you would for row-organized tables.
DB2 10.5 introduces an option for specifying not enforced (informational) primary key
constraints or unique constraints. Because the NOT ENFORCED clause does not
guarantee uniqueness, use this option only when you are certain that the data is unique.
However, not enforced primary key constraints or unique constraints require
significantly less storage and less time to create, because no internal B-tree structure is
involved. Informational primary key, foreign key, and check constraints can be very
beneficial to the query optimizer. If your data has been cleansed through the use of
rigorous extract, transform, and load (ETL) procedures, we recommend that you take
advantage of informational constraints, especially primary key constraints.
An enforced primary key constraint consumes space because a unique index is required to
enforce the constraint. Like any index, this unique index increases the cost of insert,
update, or delete operations. Verify that your application would benefit from and
requires enforcement of primary key or unique constraints.
Synopsis tables and data skipping
To improve scan performance, BLU Acceleration uses data skipping: skipping over data
that is of no interest to a query. For example, suppose a query calculates the sum of last
month’s sales. There is no need to look at data that is older than the last month. With data
skipping, BLU Acceleration can automatically skip over data that does not apply.
Internally, each column-organized user table has a companion metadata table, called a
synopsis table. The synopsis table contains all of the user table’s non-character columns
(that is, datetime, Boolean, or numeric columns) and those character columns that are
part of a primary key or foreign key definition. The synopsis table stores the minimum
and maximum values for each column across a range of rows (identified by tuple
sequence number 6 , TSN) and uses those values for data skipping during query
6
The TSN indicates which column values belong together as a logical row.
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 13 of 27
processing. A synopsis table is approximately 0.1% of the size of its user table. There is 1
row in the synopsis table for every 1024 rows in the user table.
Figure 1. A synopsis table and its relationship to a column-organized user table
You can determine the name of the synopsis table for a particular user table by querying
the SYSCAT.TABLES catalog view, as shown in the following example:
SELECT tabschema, tabname, tableorg
FROM syscat.tables
WHERE tableorg = 'C';
TABSCHEMA
-----------DB2INST1
SYSIBM
TABNAME
TABLEORG
------------------------------ -------SALES_COL
C
SYN130330165216275152_SALES_COL
C
2 record(s) selected.
It is important to understand how synopsis tables are used to optimize data skipping
during the processing of specific queries. As mentioned, the synopsis table stores
minimum and maximum values for all numeric, datetime, primary key, and foreign key
columns. If you have a column that stores dates, it is best to use the DATE data type
rather than character values so that the synopsis table enables effective data skipping.
Loading data
Column-organized tables are typically loaded from input data sources that are organized
in row format. The data is compressed and converted into column-organized format
during load or insert operations. A column-organized table is typically smaller than a
comparable row-organized table because of better compression and less space required
for auxiliary structures such as indexes or aggregates.
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 14 of 27
BLU Acceleration uses approximate Huffman encoding, prefix compression, and offset
compression. Approximate Huffman encoding is a frequency-based compression that
uses the fewest number of bits to represent the most common values. In other words, the
most common values compress the most.
BLU Acceleration builds column compression dictionaries as part of the initial load
operation on a column-organized table. Additional page-level dictionaries might be
created to take advantage of local data clustering at the page level and to further
compress the data. To optimize the compression dictionaries and ensure optimal
compression of the data, ensure that the initial load (or dictionary creation) operation
uses a representative sample of the data. However, new values that are not covered by
column compression dictionaries can still be compressed by page-level dictionaries.
Although input data parsing, key options for the LOAD command, and general load
semantics for column-organized tables and row-organized tables are the same, there are
a few key differences:

Loading data into column-organized tables has a new ANALYZE phase.

Data is converted from row-organized input into fully formatted and
compressed column-organized pages.

The synopsis tables are maintained.
Table 4 outlines the activities that occur during several phases of a load operation into a
column-organized table.
Table 4. Phases of the load process for column-organized tables
Load phase
Activity
ANALYZE
This phase occurs only if dictionaries are needed;
histograms are built to track the frequency of data values
in all columns, and column compression dictionaries are
built.
LOAD

Column and page compression dictionaries are used
to compress the data.

Compressed values are written to data pages.

Synopsis tables are maintained.
BUILD
Any unique indexes for enforced primary key
constraints are built. Although this phase always occurs
for column-organized tables, in the case of roworganized tables, it depends on the existence of indexes.
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 15 of 27
A load operation into a column-organized table automatically collects statistics according
to the profile that is defined for the table. If a profile does not exist, default automatic
statistics (equivalent to the WITH DISTRIBUTION AND SAMPLED DETAILED
INDEXES ALL clause) are collected by the RUNSTATS utility.
For optimal compression and load performance, there are two important considerations.
First, the UTIL_HEAP_SZ database configuration parameter directly affects how many
different values can be maintained during the compression dictionary build process. For
example, if a histogram that uses memory from the utility heap can only contain 1000
different repeating column values, only those 1000 repeating values can be considered
for compression, even though there could be thousands of different repeating values in a
column, which could result in less efficient compression. It is crucial to review the
UTIL_HEAP_SZ value. If the UTIL_HEAP_SZ parameter is set to AUTOMATIC, ensure
that the database memory is large enough so that there is ample memory available for
this parameter. Otherwise, update the UTIL_HEAP_SZ value to be as large as possible
prior to a first load operation into a column-organized table (which builds the histograms
and compression dictionary), and then reduce the value after the load operation
completes. UTIL_HEAP_SZ can be updated dynamically.
Second, if possible, presort the data by columns that are frequently referenced by
predicates that filter the fact table, or columns that are used to join highly filtered
dimension tables. This type of presorting can improve compression ratios and improves
the likelihood of data skipping (and better performance) by ensuring that a particular
column value is clustered on fewer data pages.
To reduce the total time that is required to load large input data files into a columnorganized table for the first time, you can reduce the duration of the ANALYZE phase by
creating a column compression dictionary with only a portion of the input data file. This
approach is outlined as follows:
1.
Create the table. Do not create any primary keys at this time.
2.
Obtain a subset of representative data from all of the data files that will be
loaded into the database. If the data is already in a DB2 database (perhaps in a
row-organized table), use a cursor with a sampling SELECT statement to obtain
the sample.
3.
Load the sample for the sole purpose of building the column compression
dictionary, as in the following example:
LOAD FROM subset_data.csv OF DEL
REPLACE RESETDICTIONARYONLY INTO mytable;
4.
Load and compress the full set of source data by using this prebuilt column
compression dictionary, as in the following example:
LOAD FROM all_data.csv OF DEL
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 16 of 27
INSERT INTO mytable;
Step 4 does not trigger the ANALYZE phase. Repeat the LOAD command until
all of the data has been loaded.
5.
Create enforced or informational primary key constraints and informational
foreign key constraints.
6.
Issue the RUNSTATS command to ensure a complete view of the table that was
populated through this multistep process.
This approach reduces the total load time but can also reduce the compression efficiency
if the sampled data is not a representative subset of the entire data set.
The new PCTENCODED column in the SYSCAT.COLUMNS catalog view represents the
percentage of values that are encoded as a result of compression for a column in a
column-organized table. If the overall compression ratio for your column-organized table
is too low, check this statistic to see if values in specific columns were left uncompressed.
If you see many columns with a very low value (or even 0) for PCTENCODED, the utility
heap might have been too small when the column compression dictionaries were created.
You might also see very low values for columns that were incrementally loaded with
data that was outside of the scope of the column compression dictionaries.
Query execution plans and the new CTQ operator
Access methods and join enumeration are greatly simplified with BLU Acceleration,
because BLU Acceleration generally considers table scans, hash joins, and hash
aggregation. Join order is still dictated by cost, so cardinality estimation is still important.
Collecting column group statistics, for example, is helpful for both column-organized
and row-organized tables.
Defining informational primary key constraints and unique constraints whenever
possible is an excellent strategy for providing the optimizer with additional information
that would not be available otherwise. This information is used to obtain good query
execution plans for BLU-accelerated tables.
The new CTQ operator for query execution plans is used to indicate the transition
between column-organized data processing (BLU Acceleration) and row-organized data
processing. All operators that appear below the CTQ operator in an execution plan are
optimized for column-organized tables. In good execution plans for column-organized
tables, the majority of operators are below the CTQ operator, and only a few rows flow
through the CTQ operator (Figure 3).
Figure 3. An example of a good query execution plan for column-organized tables
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 17 of 27
There is comprehensive support for explain in DB2 10.5 with BLU Acceleration, not just
with the engine tools (db2expln and db2exfmt), but also with a new version of the
Optim Performance Manager and the Optim Query Workload Tuner.
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 18 of 27
Database maintenance
Automatic statistics collection is enabled by default for a database that contains either
row-organized or column-organized tables. Because automatic statistics collection is also
enabled by default during a load operation into a column-organized table, an explicit
RUNSTATS operation might not be necessary, particularly in the case of single-step load
operations.
Space is automatically reclaimed for column-organized tables when
DB2_WORKLOAD=ANALYTICS or if the AUTO_REORG database configuration
parameter is set to ON. If neither of these conditions is true, you can use a REORG
TABLE…RECLAIM EXTENTS command to reclaim free space. To determine whether
there is sufficient space to reclaim, you can query the RECLAIMABLE_SPACE column in
the ADMINTABINFO administrative view. The following sample SQL returns the
amount of reclaimable space for the column-organized table named MYTABLE:
SELECT SUBSTR(a.tabname,1,20) AS tabname, RECLAIMABLE_SPACE
FROM SYSIBMADM.ADMINTABINFO
WHERE tabname LIKE 'mytable%'
ORDER BY tabname WITH UR
All other maintenance activities that involve column-organized tables, such as backup
and restore, are identical to those that involve row-organized tables.
Concurrency control (workload management)
DB2 10.5 automatically controls the resource consumption of running queries. Although
a large number of queries can be submitted concurrently, the number of queries that are
allowed to execute at one time is limited to optimize the amount of memory and CPU
resource that are available on the server for each running query.
DB2 10.5 introduces a new service subclass named
SYSDEFAULTMANAGEDSUBCLASS. Any READ DML (a work type for work classes)
queries whose estimated cost is greater than the default timeron threshold of 150000 run
in the SYSDEFAULTMANAGEDSUBCLASS instead of the SYSDEFAULTSUBCLASS.
When DB2_WORKLOAD=ANALYTICS, the default concurrency threshold on the
SYSDEFAULTMANAGEDSUBCLASS service subclass is enabled by default. The default
concurrency threshold allows a fixed number of queries into the system at a time,
specifically those queries whose estimated cost is greater than the default timeron
threshold. The fixed number is based on a heuristic calculation that factors in system
hardware attributes such as the number of CPU sockets, CPU cores, and threads per
core. The limit is calculated based on the server configuration at the time of the CREATE
DATABASE command. If changes are made to the server configuration, the concurrency
threshold can be recalculated by running the AUTOCONFIGURE command against the
database, as outlined in the following steps:
1.
Issue the AUTOCONFIGURE command:
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 19 of 27
AUTOCONFIGURE APPLY NONE
2.
Review the output, in particular the “Current and Recommended Values for
System WLM Objects” section.
Current and Recommended Values for System WLM Objects
Description
Current Value Recommended Value
----------------------------------------------------------------------------Work Action SYSMAPMANAGEDQUERIES Enabled
= Y
Y
Work Action Set SYSDEFAULTUSERWAS Enabled = Y
Y
Work Class SYSMANAGEDQUERIES Timeroncost
= 1.50000E+05
1.50000E+05
Threshold SYSDEFAULTCONCURRENT Enabled
= Y
Y
Threshold SYSDEFAULTCONCURRENT Maxvalue
= 16
18
Use the recommended value for “Threshold SYSDEFAULTCONCURRENT
Maxvalue” in place of maxvalue in the following SQL statement.
3.
Issue the following ALTER THRESHOLD statement:
ALTER THRESHOLD SYSDEFAULTCONCURRENT
WHEN CONCURRENTDBCOORDACTIVITIES > maxvalue
AND QUEUEDACTIVITIES UNBOUNDED
STOP EXECUTION
If necessary, you can define additional DB2 workload management controls to meet the
service-level agreement (SLA) requirements of the workload.
Monitoring
DB2 10.5 introduces new monitor elements for column-organized tables. When you
monitor database performance in a BLU-accelerated environment, you are seeking the
answers to several key questions. The following sections address each of these questions
and make recommendations about which monitor elements are relevant.
Is the sort heap adequate for good performance?
Review the monitor elements in Table 5 to assess how the sort heap is being used.
Table 5. Monitor elements that are relevant to column-organized tables, grouped by the
areas that they monitor
Area
Elements
Sort heap
SORT_HEAP_ALLOCATED,
SORT_SHRHEAP_ALLOCATED,
SORT_SHRHEAP_TOP
Sort
ACTIVE_SORTS, SORT_OVERFLOWS
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 20 of 27
Join
ACTIVE_HASH_JOINS, TOTAL_HASH_JOINS,
HASH_JOIN_OVERFLOWS,
HASH_JOIN_SMALL_OVERFLOWS,
POST_SHRTHRESHOLD_HASH_JOINS,
POST_THRESHOLD_HASH_JOINS
Group by
ACTIVE_HASH_GRPBYS,
TOTAL_HASH_GRPBYS,
HASH_GRPBY_OVERFLOWS,
POST_THRESHOLD_HASH_GRPBYS
“Overflows” indicate that a query could not obtain all of the sort memory that it needed,
which can be the result of a SORTHEAP setting that is too small, or a result of attaining
the SHEAPTHRES_SHR threshold because the SORTHEAP setting is too large for the
level of concurrency in the workload. Review the value that is returned by the
SORT_SHRHEAP_TOP monitor element and the SHEAPTHRES_SHR database
configuration parameter setting to determine which case is applicable.
“Post threshold sorts” indicate sorts that are throttled below the SORTHEAP setting.
“Small overflows” indicate an overflow of 10% or less.
Overflows or post threshold sorts might indicate the need to increase the SORTHEAP
setting to reduce spills and improve performance, or the need to reduce the SORTHEAP
setting relative to the SHEAPTHRES_SHR setting because of higher concurrency in the
workload. In general, it is typically better to increase SHEAPTHRES_SHR rather than
SORTHEAP, in order to improve overall workload performance.
Is the table suitably organized?
By dividing the NUM_COLUMNS_REFERENCED monitor element by the
SECTION_EXEC_WITH_COL_REFERENCES monitor element, you can determine the
average number of columns that a particular query accesses. If this average is much less
than the number of columns in the table, the workload favors column organization.
How well are queries performing?
New monitor elements were added to measure specific aspects of column-organized data
processing. These measurements include elapsed time (reported by the
TOTAL_COL_TIME element), processing time excluding lock wait times, I/O, and other
items (reported by the TOTAL_COL_PROC_TIME element), and the number of times the
CTQ operator is accessed (reported by the TOTAL_COL_EXECUTIONS element).
You can use these elements in the context of other “time-spent” monitor elements to
determine how much time was spent, per thread, performing column-organized data
processing. For example, if you want to know what portion of a query was executed in
column-organized form and what portion was executed in row-organized form, you can
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 21 of 27
compare TOTAL_COL_TIME to TOTAL_SECTION_TIME. A large ratio suggests an
optimal execution plan for BLU Acceleration.
For more information about the time-spent monitor elements, see
http://pic.dhe.ibm.com/infocenter/db2luw/v10r5/index.jsp?topic=%2Fcom.ibm.db2.luw.a
dmin.mon.doc%2Fdoc%2Fc0056890.html.
How is the buffer pool performing?
New versions of existing monitor elements (for example, COL_HIT_RATIO_PERCENT
and COL_PHYSICAL_READS) were added to enable you to monitor buffer pool usage
by column-organized tables.
How are the prefetchers performing?
New versions of existing monitor elements (for example,
POOL_QUEUED_ASYNC_COL_PAGES and SKIPPED_PREFETCH_COL_P_READS)
were added to enable you to monitor prefetching for column-organized tables.
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 22 of 27
Conclusion
The best practices that are presented in this paper are intended to get you up and
running with BLU Acceleration, a new technology for analytic queries that is introduced
in DB2 for Linux, UNIX, and Windows Version 10.5.
The paper began with an overview of the cornerstones of this technology, including
column-organized tables, actionable compression, parallel vector processing, and data
skipping.
This was followed by recommendations pertaining to hardware and software selection,
guidelines for identifying the optimal workloads for BLU Acceleration, and information
about capacity planning, memory, and I/O.
A section on system configuration included information on how to have DB2
automatically set up an optimal default configuration for analytic workloads with one
easy step, plus information on how to derive the greatest benefits from the memory that
is available on your system.
The remaining sections told you how to implement and best use DB2 with BLU
Acceleration. We included the most important information about column-organized
tables and told you how to ensure that synopsis tables are performing optimally. These
sections also gave you best practices for loading data, and information that will help you
to keep your system running at top performance levels.
By leveraging these best practices, you can use DB2 10.5 with BLU Acceleration to gain
benefits for your enterprise by performing analytic processing that is super fast and super
easy – just load and go!
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 23 of 27
Further reading

Information Management best practices:
http://www.ibm.com/developerworks/data/bestpractices/

DB2 for Linux, UNIX, and Windows best practices:
http://www.ibm.com/developerworks/data/bestpractices/db2luw/

DB2 with BLU Acceleration: http://www-01.ibm.com/software/data/db2/linuxunix-windows/db2-blu-acceleration/

“Super Analytics, Super Easy - Introducing IBM DB2 10.5 with BLU Acceleration”:
http://ibmdatamag.com/2013/04/super-analytics-super-easy/

DB2 10.5 Information Center:
http://pic.dhe.ibm.com/infocenter/db2luw/v10r5/index.jsp
Contributors
Significant contributions were made to the content of this paper by the following
persons:
Chris Eaton
WW IM Technical Sales and Competitive
Specialist, IBM Canada Lab
David Kalmuk
Architect, DB2 LUW Workload
Management, Monitoring, Process Model,
IBM Canada Lab
Calisto Zuzarte
Senior Technical Staff Member, IBM
Canada Lab
Acknowledgements
We thank the following people whose valuable input helped to improve the quality of
this paper:
Gopi Attaluri
Serge Boivin
Eric Koeck
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 24 of 27
Nela Krawez
Christina Lee
Sam Lightstone
Mathias Nicola
Huaxin Zhang
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 25 of 27
Notices
This information was developed for products and services offered in the U.S.A.
IBM may not offer the products, services, or features discussed in this document in other
countries. Consult your local IBM representative for information on the products and services
currently available in your area. Any reference to an IBM product, program, or service is not
intended to state or imply that only that IBM product, program, or service may be used. Any
functionally equivalent product, program, or service that does not infringe any IBM
intellectual property right may be used instead. However, it is the user's responsibility to
evaluate and verify the operation of any non-IBM product, program, or service.
IBM may have patents or pending patent applications covering subject matter described in
this document. The furnishing of this document does not grant you any license to these
patents. You can send license inquiries, in writing, to:
IBM Director of Licensing
IBM Corporation
North Castle Drive
Armonk, NY 10504-1785
U.S.A.
The following paragraph does not apply to the United Kingdom or any other country where
such provisions are inconsistent with local law: INTERNATIONAL BUSINESS MACHINES
CORPORATION PROVIDES THIS PUBLICATION "AS IS" WITHOUT WARRANTY OF ANY KIND, EITHER
EXPRESS OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF NONINFRINGEMENT, MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. Some states do
not allow disclaimer of express or implied warranties in certain transactions, therefore, this
statement may not apply to you.
Without limiting the above disclaimers, IBM provides no representations or warranties
regarding the accuracy, reliability or serviceability of any information or recommendations
provided in this publication, or with respect to any results that may be obtained by the use of
the information or observance of any recommendations provided herein. The information
contained in this document has not been submitted to any formal IBM test and is distributed
AS IS. The use of this information or the implementation of any recommendations or
techniques herein is a customer responsibility and depends on the customer’s ability to
evaluate and integrate them into the customer’s operational environment. While each item
may have been reviewed by IBM for accuracy in a specific situation, there is no guarantee
that the same or similar results will be obtained elsewhere. Anyone attempting to adapt
these techniques to their own environment does so at their own risk.
This document and the information contained herein may be used solely in connection with
the IBM products discussed in this document.
This information could include technical inaccuracies or typographical errors. Changes are
periodically made to the information herein; these changes will be incorporated in new
editions of the publication. IBM may make improvements and/or changes in the product(s)
and/or the program(s) described in this publication at any time without notice.
Any references in this information to non-IBM websites are provided for convenience only
and do not in any manner serve as an endorsement of those websites. The materials at those
websites are not part of the materials for this IBM product and use of those websites is at your
own risk.
IBM may use or distribute any of the information you supply in any way it believes
appropriate without incurring any obligation to you.
Any performance data contained herein was determined in a controlled environment.
Therefore, the results obtained in other operating environments may vary significantly. Some
measurements may have been made on development-level systems and there is no
guarantee that these measurements will be the same on generally available systems.
Furthermore, some measurements may have been estimated through extrapolation. Actual
results may vary. Users of this document should verify the applicable data for their specific
environment.
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 26 of 27
Information concerning non-IBM products was obtained from the suppliers of those products,
their published announcements or other publicly available sources. IBM has not tested those
products and cannot confirm the accuracy of performance, compatibility or any other
claims related to non-IBM products. Questions on the capabilities of non-IBM products should
be addressed to the suppliers of those products.
All statements regarding IBM's future direction or intent are subject to change or withdrawal
without notice, and represent goals and objectives only.
This information contains examples of data and reports used in daily business operations. To
illustrate them as completely as possible, the examples include the names of individuals,
companies, brands, and products. All of these names are fictitious and any similarity to the
names and addresses used by an actual business enterprise is entirely coincidental.
COPYRIGHT LICENSE: © Copyright IBM Corporation 2013. All Rights Reserved.
This information contains sample application programs in source language, which illustrate
programming techniques on various operating platforms. You may copy, modify, and
distribute these sample programs in any form without payment to IBM, for the purposes of
developing, using, marketing or distributing application programs conforming to the
application programming interface for the operating platform for which the sample
programs are written. These examples have not been thoroughly tested under all conditions.
IBM, therefore, cannot guarantee or imply reliability, serviceability, or function of these
programs.
Trademarks
IBM, the IBM logo, and ibm.com are trademarks or registered trademarks of International
Business Machines Corporation in the United States, other countries, or both. If these and
other IBM trademarked terms are marked on their first occurrence in this information with a
trademark symbol (® or ™), these symbols indicate U.S. registered or common law
trademarks owned by IBM at the time this information was published. Such trademarks may
also be registered or common law trademarks in other countries. A current list of IBM
trademarks is available on the Web at “Copyright and trademark information” at
www.ibm.com/legal/copytrade.shtml
Windows is a trademark of Microsoft Corporation in the United States, other countries, or
both.
UNIX is a registered trademark of The Open Group in the United States and other countries.
Linux is a registered trademark of Linus Torvalds in the United States, other countries, or both.
Other company, product, or service names may be trademarks or service marks of others.
Contacting IBM
To provide feedback about this paper, write to db2docs@ca.ibm.com
To contact IBM in your country or region, check the IBM Directory of Worldwide
Contacts at http://www.ibm.com/planetwide
To learn more about IBM Information Management products, go to
http://www.ibm.com/software/data/
Optimizing analytic workloads using DB2 10.5 with BLU Acceleration
Page 27 of 27