© Peak Indicators Limited OBIEE and Database Performance Tuning Antony Heljula Technical Architect
Dec 25, 2014
© Peak Indicators Limited
OBIEE and Database Performance Tuning
Antony Heljula
Technical Architect
© Peak Indicators Limited 2
Agenda
Aim of Presentation
Test Scenario
Performance Tests
Summary & Conclusion
And Finally….
© Peak Indicators Limited 3
Aim of Presentation
© Peak Indicators Limited 4
OBIEE Performance Issues
When reports take a long time to return results, end users tend to say “OBIEE does not perform well”
The truth is that most (99%?) of the processing typically takes place on the underlying data sources If another BI tool was used to deliver the
same reports, the queries would probably still run just as slowly
It would therefore be more accurate to say that it is the databases that are not performing well!
WebLogic
BI Server
BI Presentation Services
“Analytics” BI Plug-In
© Peak Indicators Limited 5
Aim of Presentation
We are going to investigate how a set of OBIEE Dashboard queries can be tuned to deliver satisfactory performance
We’ll begin testing with a set of dashboards that perform very poorly
We’ll then implement a number of tuning features one-by-one to see how things improve….hopefully by the end of it we’ll have decent performance!
The aim will be to get all reports to return in less than 10 seconds
The performance test results will be captured and reported on using Usage Tracking
© Peak Indicators Limited 6
Aim of Presentation
We are going to assume that the database has an optimum and balanced configuration, so that any performance issues are “software” related and not “hardware”
© Peak Indicators Limited 7
Aim of Presentation
On an Oracle database, there are many known ways to improve query performance. For example:
Gather statistics
Remove snow-flakes
Star Transformation
Partitioning
Bitmap Indexes
We will attempt to answer the following questions:
Do they actually work?
What performance gains do they actually deliver?
In what situations are they most effective?
Bitmap Join Indexes Compression Parallel Query Aggregation Denormalization
© Peak Indicators Limited 8
Aim of Presentation
The aim is to deliver satisfactory performance using standard relational database features
The customer won’t be happy if they are told mid-way through UAT that they need to purchase additional software licenses e.g. Oracle OLAP / Essbase
We will however be looking at “Partitioning”, although this option does require additional license cost it is generally purchased by most/all customers who have large data volumes
We are going to avoid the use of “hints”
Hints effectively hard-code the optimization rules for database queries
Hints are “old” technology dating back to the Rule Base Optimizer (RBO), they do not take into account the size and complexity of the query in the way that the Cost Based Optimizer (CBO) does
Hints should always be the last resort (in my view)
Notes
© Peak Indicators Limited 9
Test Scenario
© Peak Indicators Limited 10
Test Scenario
Dell Latitude E6400:
Windows 7 64-bit
2.54Ghz dual-core CPU
8GB RAM
250GB SATA internal hard disk (7,200 RPM)
Software:
Oracle Database Enterprise Edition 11g R2
Oracle BI Enterprise Edition 11.1.1.3
Hardware
© Peak Indicators Limited 11
Test Scenario
To conduct the investigation, a database data-model was built entirely from scratch:
Data-Model
Fact (Daily Summary) Time Dimension
Product Dimension
Customer Dimension
Organization Dimension
© Peak Indicators Limited 12
Test Scenario
The tables were then populated with a completely fabricated set of data
Number of records in each table is show in red
Approximately 10GB total volume
Data Volumes
30 Million 11,000
5,000
500
500,000
1,000
5,000
10,000
9,000
© Peak Indicators Limited 13
Test Scenario
Examples of the fabricated data that was generated:
Fabricated Data
© Peak Indicators Limited 14
Test Scenario
An RPD was developed using all the modern best-practices for a star-schema data-model:
RPD
© Peak Indicators Limited 15
Test Scenario
A “Sales Orders” dashboard was created containing 6 pages with only 1 analysis per page
4 pages contained a “summary” analysis (month level or above)
2 pages contained a “detail” analysis (day/week level)
Dashboards
© Peak Indicators Limited 16
Test Scenario
Each performance test was conducted manually but in a strict sequence:
1. Log on
2. Go to each dashboard page one-by-one (in the same order every time)
3. Only move to the next page once the current page has returned results
4. Log off
To ensure every performance test was fair, the following steps were taken before each test was conducted:
Restart database (to purge all database cache)
Purge BI Server cache
Purge BI Presentation Services cache
Test Rules
© Peak Indicators Limited 17
Test Scenario
Before starting the tests, we had no indication as to what the results would be – there was no certainty any firm conclusions could be made afterwards
No attempt was made to “prepare” the data-model or the performance tests so that the final results would look good or bad
When each feature was tested, no effort was made to tune the particular feature – we simply used the default settings to see if it would work straight “out of the box”
If we ran the exact same test twice, the timings could vary by about 5-10 seconds or 10%. We will therefore assume that timings have to be different by 10% or >10 seconds in order to conclude that any tuning feature has made a difference
Final Notes
© Peak Indicators Limited 18
Performance Tests
© Peak Indicators Limited 19
1. Starting Point
To begin with, the data-model had the following features / issues:
Plenty of snow-flaking
No statistics generated (RBO is therefore in use)
B-Tree indexes used throughout
Star Transformation disabled
Overview
© Peak Indicators Limited 20
1. Starting Point
446 seconds in total to run all 6 dashboard pages in sequence
The 4 “summary” reports all perform poorly
The 2 “detail” reports are returning in less than 10 seconds
Result
© Peak Indicators Limited 21
2. Gather Stats 30%
If you have a performance issue with a database query, one of the first questions you will get asked is “have you analysed your tables and indexes?”
The Oracle Database comes with a “Cost Based Optimizer” (CBO) which determines the most appropriate way to process your query based on the size and contents of the required tables and their indexes But you need to analyze your tables (or “gather statistics”) in order for the
CBO to be used, otherwise the “Rule Based Optimizer” (RBO) is used
With large data warehouses, it is sometimes not possible to analyze all the data and indexes as the process will take too long, so this first test will see if gathering statistics using only a 30% sample of data is sufficient to make the CBO work efficiently
EXEC DBMS_STATS.GATHER_TABLE_STATS (ownname => 'PEAKTEST',
tabname => 'WH_CUSTOMER',
estimate_percent => 30);
Overview
© Peak Indicators Limited 22
2. Gather Stats 30%
Result
© Peak Indicators Limited 23
2. Gather Stats 30%
Surprisingly, performance got worse overall by 52%!
“Summary” Reports
2 reports had significant improvements >65%
2 reports did not change more than 10%
“Detail” Reports
Both detail reports suffered worse performance, one of them actually took 125 times longer than before!
The over-use of “hash joins” is causing the issue
Summary
© Peak Indicators Limited 24
3. Gather Stats 100%
If you have performance issues after gathering statistics using a 30% sample of data, it is quite likely that Oracle Support or a DBA will recommend that you try gathering statistics using a greater sample of data
So this next test will provide an indication as to whether gathering statistics using a “full” 100% sample will make a significant difference…
EXEC DBMS_STATS.GATHER_TABLE_STATS (ownname => 'PEAKTEST',
tabname => 'WH_CUSTOMER',
estimate_percent => 100);
Overview
© Peak Indicators Limited 25
2. Gather Stats 30%
Result
© Peak Indicators Limited 26
3. Gather Stats 100%
Performance improved overall by 9%
“Summary” Reports
Performance was essentially the same as with a 30% statistics sample
Performance levels are still far from acceptable
“Detail” Reports
Performance did improve overall by 14%
Performance levels are still far from acceptable
Summary
© Peak Indicators Limited 27
4. Star Transformation (No Bitmaps)
The Oracle database has a special tuning feature designed for optimising “star schemas” queries, it is enabled by setting the following parameter:
STAR_TRANSFORMATION_ENABLED = TRUE
It is however often overlooked that the documentation recommends that, for this feature to work properly, all the foreign key columns on your fact tables should have “bitmap indexes” created and not just “b-tree indexes”:
This test will see what the effect is if you don’t have bitmap indexes on your foreign key columns…
Overview
© Peak Indicators Limited 28
4. Star Transformation (No Bitmaps)
Result
© Peak Indicators Limited 29
4. Star Transformation (No Bitmaps)
Performance overall got slightly worse, but not by much
“Summary” Reports
Enabling Star Transformation without bitmap indexes had no real effect
“Detail” Reports
1 report improved by 50% (11 seconds) but overall there was no significant difference
Summary
© Peak Indicators Limited 30
5. Star Transformation (Bitmaps)
This test will see what the impact is when Star Transformation is enabled with bitmap indexes created for the foreign keys on the fact table
(as recommended by Oracle)
CREATE BITMAP INDEX WH_SALES_FACT_DAY_XFK_1 ON WH_SALES_FACT_DAY_XFK (CUSTOMER_KEY);
CREATE BITMAP INDEX WH_SALES_FACT_DAY_XFK_2 ON WH_SALES_FACT_DAY_XFK (PRODUCT_KEY);
CREATE BITMAP INDEX WH_SALES_FACT_DAY_XFK_3 ON WH_SALES_FACT_DAY_XFK (ORG_KEY);
CREATE BITMAP INDEX WH_SALES_FACT_DAY_XFK_4 ON WH_SALES_FACT_DAY_XFK (TIME_DAY_KEY);
Overview
© Peak Indicators Limited 31
5. Star Transformation (Bitmaps)
Result
© Peak Indicators Limited 32
5. Star Transformation (Bitmaps)
Performance improved overall by 50%!
Finally things are starting to get under control, but we still need to do better
“Summary” Reports
15% overall improvement
2 of the reports had significant improvements
“Detail” Reports
A 78% improvement overall, although one of the reports increased from 8 to 60 seconds
Summary
© Peak Indicators Limited 33
6. Remove Snow-Flakes (Dim Cols.)
Snow-flaking occurs when you have a chain of Dimension tables joined together
The Oracle Data-Warehousing states that performance can be improved (especially with “Star Transformation”) if you data-model does not consist of snow-flakes
So this test will show what happens when we eliminate snow-flaking by combining the snow-flaked tables into the main dimension table (forming a pure star-schema):
Overview
© Peak Indicators Limited 34
6. Remove Snow-Flakes (Dim Cols.)
Result
© Peak Indicators Limited 35
6. Remove Snow-Flakes (Dim Cols.)
Surprisingly, it made things slightly worse, no reports showed improvement!
“Summary” Reports
Moving snow-flaked dimension columns into the main dimension table had no real effect, just a few seconds worse overall
“Detail” Reports
Moving snow-flaked dimension columns into the main dimension table had no real effect, just a few seconds worse overall
Summary
© Peak Indicators Limited 36
7. Remove Snow-Flakes (Add FKs)
As the previous method of eliminating snow-flakes did not work, this time we will try a slightly different approach
In this test we will eliminate snow-flakes by adding extra foreign keys to the central fact table which join directly to the snow-flaked dimension tables (with “bitmap indexes” created)
Will this more perfect “star-schema” improve performance significantly?
Overview
© Peak Indicators Limited 37
7. Remove Snow-Flakes (Add FKs)
Result
* Test 6 was backed out
© Peak Indicators Limited 38
7. Remove Snow-Flakes (Add FKs)
It made things much worse….response times more than doubled!
Adding more FKs made the fact table much larger, making it longer to scan?
More FKs more complexity?
“Summary” Reports
136% worse overall
“Detail” Reports
90% worse overall
Summary
© Peak Indicators Limited 39
8. Bitmap Join Indexes
Oracle promote “bitmap join indexes” as a way of increasing performance by “orders of magnitude”
A bitmap join index is a bitmap index which stores the actual results of a join between two or more tables. For example, here is a bitmap join index that stores the result of the joins from the fact table to four columns in the “Day” dimension table:
CREATE BITMAP INDEX WH_SALES_FACT_DAY_BMJ_8
ON WH_SALES_FACT_DAY (PER_NAME_YEAR, PER_NAME_QTR, PER_NAME_MONTH,PER_NAME_WEEK)
FROM WH_SALES_FACT_DAY F, WH_TIME_DAY TD
WHERE F.TIME_DAY_KEY = TD.TIME_DAY_KEY;
For this test 8 bitmap join indexes were created to store the join results of all the joins made between the fact and dimension tables across the 6 dashboard pages
Overview
© Peak Indicators Limited 40
8. Bitmap Join Indexes
Result
* Tests 6 and 7 were backed out
© Peak Indicators Limited 41
8. Bitmap Join Indexes
Bitmap join indexes improved performance overall by 55 seconds (18%)
However only 1 “detail” report actually used a bitmap join index It seems the Oracle database will only use a bitmap join index if it deems the query is
at a low enough granularity to make it worthwhile
“Summary” Reports No changes in performance Bitmap join indexes did not get used for any of the summary reports!
“Detail” Reports One report reduced from 60 down to only 3 seconds. A massive improvement! The other detail report did not change, bitmap join indexes did not get used
Summary
© Peak Indicators Limited 42
8. Bitmap Join Indexes
Bitmap Join indexes do have some restrictions:
Overview
© Peak Indicators Limited 43
9. Partition by Month
Table “partitioning” is a very common recommendation when dealing with very large tables. On a data warehouse the most obvious thing to do is partition your large fact tables by Day / Week / Month
It is essential to make sure that your partition strategy will result in effect “partition pruning”. This means that, when OBIEE runs a query, the database will know it only needs to scan a sub-set of the table partitions, rather than scanning the whole table
In our case, we test out partitioning tables by “Month” based on the existing “TIME_DAY_KEY” foreign key column on the fact table, it is in YYYYMMDD format
When OBIEE queries for a specific time period, the database will lookup the TIME_DAY_KEY values in the “Time” dimension table and know exactly which fact table partitions to scan:
Overview
TIME_DAY_KEY
TIME_DAY_KEY
© Peak Indicators Limited 44
9. Partition by Month
Here is the SQL statement used to build our partitioned fact table which has a partition for each of the 24 months of data that exist:
Overview
© Peak Indicators Limited 45
9. Partition by Month
Note that when you partition a table, you need to create “LOCAL” bitmap indexes on the foreign key columns
“LOCAL” means that the bitmap index will also be broken down into partitions, potentially improving performance even further as the smaller index partitions will not take so long to process:
CREATE BITMAP INDEX WH_SALES_FACT_DAY_1 ON WH_SALES_FACT_DAY (CUSTOMER_KEY) LOCAL;
CREATE BITMAP INDEX WH_SALES_FACT_DAY_2 ON WH_SALES_FACT_DAY (PRODUCT_KEY) LOCAL;
CREATE BITMAP INDEX WH_SALES_FACT_DAY_3 ON WH_SALES_FACT_DAY (ORG_KEY) LOCAL;
CREATE BITMAP INDEX WH_SALES_FACT_DAY_4 ON WH_SALES_FACT_DAY (TIME_DAY_KEY) LOCAL;
Remember to analyze / gather statistics afterwards!
Overview
© Peak Indicators Limited 46
9. Partition by Month
Result
* Tests 6 and 7 were backed out
© Peak Indicators Limited 47
9. Partition by Month
Overall performance improved by 57 seconds (23%)
“Summary” Reports
3 out of the 4 summary reports had excellent improvement >50%
1 summary report took 30 seconds (33%) longer. This report contains two “Year Ago” and “Year-to-Date” calculations which meant that most/all the table partitions had to be scanned (scanning lots of small partitions may therefore be less efficient than scanning a smaller number of larger partitions)
“Detail” Reports
A good overall improvement of 6 seconds (35%)
Summary
© Peak Indicators Limited 48
10. Parallel Query (Auto)
The Oracle database offers parallel query capability, the idea being that one sequential task can be broken down in the multiple tasks running in parallel
Parallel query is a good contender once you have implemented table partitioning, is the Oracle database can easily process each partition in parallel
On our test environment we have 1 CPU and 1 disk – not the most appropriate environment for testing parallel query
For this test, we are enabling parallel query simply by enabling the “auto tuning” parallel query feature for the Oracle Database 11g R2
This feature ignores any “parallel” degree settings you may have manually set for your tables:
alter system set PARALLEL_DEGREE_POLICY=AUTO;
Overview
© Peak Indicators Limited 49
10. Parallel Query (Auto)
Result
* Tests 6 and 7 were backed out
© Peak Indicators Limited 50
A slight performance improvement overall to the “Summary” reports, the “Detail” reports were unaffected
NOTE: Even with no parallelism enabled, our single disk was already operating at >80% capacity
So by enabling parallelism there was not much for improvement as our disk was already near to bottlenecking
It would be better to run this test on a system with >1 CPU and >1 disk
“Summary” Reports
The longest running report showed the best improvement of 30 seconds (25%)
“Detail” Reports
The detail reports had no change, the explain plans showed that the Oracle Database did not attempt to use parallel query for the granular queries (an excellent feature)
10. Parallel Query (Auto)
Summary
© Peak Indicators Limited 51
10. Parallel Query (Auto)
Explain Plans with PARALLEL_DEGREE_POLICY=AUTO
“Summary” report has parallel query
“Detail” report has no parallel query
© Peak Indicators Limited 52
11. Compression
The Oracle database allows you to store data in a compressed (zipped) format, with three possible benefits: Reduced overall storage requirements
Performance improvement due to less disk reads (each record uses less space)
Improved memory efficiency (data is stored in memory in compressed format)
The potential drawback is with higher CPU activity since all data blocks have to be uncompressed at run-time in order to be processed Compression may be a good option in our test environment, since the CPU usage is
small compared to the disk usage (20-40% versus 80-100%)
Compressing our fact table reduced the data volume from 1.46GB down to 1.09GB, a reduction of about 25% Database was given a default 8K block size, increasing to 32/64K could increase the
compression ratio significantly
NOTE: bitmap indexes always store data in compressed format
Overview
© Peak Indicators Limited 53
11. Compression
Here is the SQL statement used to build our compressed and partitioned fact table:
Overview
NOTE: There are limitations! e.g. insert records with the hint: INSERT /*+ APPEND */
© Peak Indicators Limited 54
11. Compression
Result
* Tests 6 and 7 were backed out
© Peak Indicators Limited 55
11. Compression
With a 25% compression ratio, performance improved by about 15% across the board, although it had more effect on the “Summary” reports where more data was being extracted from disk
NOTE: Oracle also has an “Advanced Compression” feature which is designed to provide even greater compression ratios
“Summary” Reports
Approximately 15% improvement
“Detail” Reports
No real change
Summary
© Peak Indicators Limited 56
Many customers can find themselves in this situation: we have already implemented several different tuning mechanisms but our “Summary” dashboards are still taking far too long…..an average of 41 seconds each (with only 1 user!)
Apart from adding more hardware (CPUs/memory/disks), what is the next option?
Sometimes there is no alternative to summarising the data in order to reduce the amount of data being processed
12. Aggregation (MVs)
Overview
© Peak Indicators Limited 57
12. Aggregation (MVs)
Implementing an aggregate strategy is often a balancing act:
You want to keep the number of aggregates to an absolute minimum
You want to summarise the data as much as possible
BUT
You want aggregates that can serve multiple reports/dashboards (plus ad hoc)
You want aggregates to support multiple drill-down levels across many hierarchies
Sometimes you have no choice other than to create an aggregate for each dashboard (in which case you may need to consider cube engines, more hardware etc):
Overview
© Peak Indicators Limited 58
12. Aggregation (MVs)
In our test environment, we have managed to build a single aggregate which can:
Reduce the number of records from 31M down to 8M
Support all 4 “Summary” dashboards
Provide at least 2 levels of drill-down across all the hierarchies
Overview
© Peak Indicators Limited 59
12. Aggregation (MVs)
HINT: Use the “Number of Elements” in the RPD to help you determine how much more efficient an aggregate table could be
Overview
© Peak Indicators Limited 60
12. Aggregation (MVs)
Aggregate tables on an Oracle database can be either:
Standard physical tables (updated or rebuilt during ETL)
Materialized Views
In theory there is no difference between the two options as they both contain snapshots of the summarised data
Materialized Views are easier to maintain and quicker to develop
If you use Materialized Views, then please note:
Do not rely too heavily on the “Query Rewrite” feature as OBIEE can generate queries that are too complex for the database to rewrite
It is often better to model the MVs into the RPD as if they were normal physical aggregate tables (so you rely on OBIEE “aggregate navigation” rather than database “query rewrite”)
Remember to partition your MVs and create bitmap indexes on any foreign keys
Overview
© Peak Indicators Limited 61
12. Aggregation (MVs)
s
Overview
MV partitioned by month
LOCAL bitmap indexes (don’t forget to analyze)
© Peak Indicators Limited 62
12. Aggregation (MVs)
Result
* Tests 6 and 7 were backed out
© Peak Indicators Limited 63
12. Aggregation (MVs)
Finally! All queries are now taken 10 seconds or less
“Summary” Reports
All reports improved, running over 7 times faster than before! (a total of 143 seconds down to 23)
“Detail” Reports
The aggregate MVs were designed for the summary reports, so the detail reports were unaffected and perform exactly the same as before
Summary
© Peak Indicators Limited 64
Summary & Conclusion
© Peak Indicators Limited 65
Summary
The biggest gains for “Detail” reports: • Star Transformation • Bitmap Join Indexes • Partitioning
The biggest gains for “Summary” reports: • Gathering Stats 30% • Aggregation • Star Transformation • Partitioning
© Peak Indicators Limited 66
Conclusion
The Oracle database provides a wide variety of tuning features, but you need to adopt a combination of tuning features
The various tuning mechanisms suit different types of reports
If you have performance issues with “Summary” reports then consider: Gathering statistics 30/100% Star Transformation Partitioning Parallel Query Aggregation Compression
If you have performance issues with “Detail” reports then consider: Gathering statistics 100% (as opposed to 30%) Star Transformation Partitioning Bitmap Join Indexes
© Peak Indicators Limited 67
Conclusion
Always test thoroughly! Whilst the various tuning mechanisms can lead to an overall positive improvement, there can be surprises where specific reports suffer worse performance (sometimes significantly worse):
© Peak Indicators Limited 68
And Finally….
What happens if none of the tuning mechanisms discussed give you the desired level of performance?
Buy bigger/better hardware (more CPUs+Disks+Memory)
Archive data to reduce volume
Oracle OLAP (now properly supported with OBIEE 11.1.1.5)
Oracle Essbase
Oracle Exadata
© Peak Indicators Limited
Questions?
© Peak Indicators Limited
Helping Your Business Intelligence Journey