Mondrian Documentation

By Sherman Wood and Julian Hyde; last updated November, 2007.

Contents

1. Introduction
2. A generalized tuning process for Mondrian
3. Recommendations for database tuning
4. Aggregate Tables, Materialized Views and Mondrian
   1. Choosing aggregate tables
5. Aggregate Generator (AggGen)
6. Optimizing Calculations with the Expression Cache

Introduction 

As with any data warehouse project, dealing with volumes is always the make or break issue. Mondrian has its own issues, based on its architecture and goals of being cross platform. Here are some experiences and comments.

From the Mondrian developer's mailing list in February, 2005 - an example of unoptimized performance:

When Mondrian initializes and starts to process the first queries, it makes SQL calls to get member lists and determine cardinality, and then to load segments into the cache. When Mondrian is closed and restarted, it has to do that work again. This can be a significant chunk of time depending on the cube size. For example in one test an 8GB cube (55M row fact table) took 15 minutes (mostly doing a group by) before it returned results from its first query, and absent any caching on the database server would take another 15 minutes if you closed it and reopened the application. Now, this cube was just one month of data; imagine the time if there was 5 years worth.

Since this time, Mondrian has been extended to use aggregate tables and materialized views, which have a lot of performance benefits that address the above issue.

From Julian:

I'm surprised that people can run 10m+ row fact tables on Mondrian at all, without using aggregate tables or materialized views.

From Sherman:

Our largest site has a cube with currently ~6M facts on a single low end Linux box running our application with Mondrian and Postgres (not an ideal configuration), without aggregate tables, and gets sub second response times for the user interface (JPivot). This was achieved by tuning the database to support the queries being executed, modifying the OS configuration to best support Postgres execution (thanks Josh!) and adding as much RAM as possible.

A generalized tuning process for Mondrian 

The process for addressing performance of Mondrian is a combination of design, hardware, database and other configuration tuning. For really large cubes, the performance issues are driven more by the hardware, operating system and database tuning than anything Mondrian can do.

• Have a reasonable physical design for requirements, such as a data warehouse and specific data marts
• Architect the application effectively
  • Separate the environment where Mondrian is executing from the DBMS
  • If possible: separate UI processing from the environment where Mondrian is caching
• Have adequate hardware for the DBMS
• Tune the operating system for the DBMS
• Add materialized views or aggregate tables to support specific MDX queries (see Aggregate Tables and AggGen below)
• Tune the DBMS for the specific SQL queries being executed: that is, indexes on both the dimensions and fact table
• Tune the Mondrian cache: the larger the better

Recommendations for database tuning 

As part of database tuning process, enable SQL tracing and tail the log file. Run some representative MDX queries and watch which SQL statements take a long time. Tune the database to fix those statements and rerun.

• Indexes on primary and foreign keys
• Consider enabling foreign keys
• Ensure that columns are marked NOT NULL where possible
• If a table has a compound primary key, experiment with indexing subsets of the columns with different leading edges. For example, for columns (a, b, c) create a unique index on (a, b, c) and non-unique indexes on (b, c) and (c, a). Oracle can use such indexes to speed up counts.
• On Oracle, consider using bitmap indexes for low-cardinality columns. (Julian implemented the Oracle's bitmap index feature, and he's rather proud of them!)
• On Oracle, Postgres and other DBMSs, analyze tables, otherwise the cost-based optimizers will not be used

Mondrian currently uses 'count(distinct ...)' queries to determine the cardinality of dimensions and levels as it starts, and for your measures that are counts, that is, aggregator="count". Indexes might speed up those queries -- although performance is likely to vary between databases, because optimizing count-distinct queries is a tricky problem.

Aggregate Tables, Materialized Views and Mondrian 

The best way to increase the performance of Mondrian is to build a set of aggregate (summary) tables that coexist with the base fact table. These aggregate tables contain pre-aggregated measures build from the fact table.

Some databases, particularly Oracle, can automatically create these aggregations through materialized views, which are tables created and synchronized from views. Otherwise, you will have to maintain the aggregation tables through your data warehouse load processes, usually by clearing them and rerunning aggregating INSERTs.

Aggregate tables are introduced in the Schema Guide, and described in more detail in their own document, Aggregate Tables.

Choosing aggregate tables 

It isn't easy to choose the right aggregate tables. For one thing, there are so many to choose from: even a modest cube with six dimensions each with three levels has 6⁴ = 1296 possible aggregate tables! And aggregate tables interfere with each other. If you add a new aggregate table, Mondrian may use an existing aggregate table less frequently.

Missing aggregate tables may not even be the problem. Choosing aggregate tables is part of a wider performance tuning process, where finding the problem is more than half of the battle. The real cause may be a missing index on your fact table, your cache isn't large enough, or (if you're running Oracle) the fact that you forgot to compute statistics. (See recommendations, above.)

Performance tuning is an iterative process. The steps are something like this:

1. Choose a few queries which are typical for those the end-users will be executing.
2. Run your set of sample queries, and note how long they take. Now the cache has been primed, run the queries again: has performance improved?
3. Is the performance good enough? If it is, stop tuning now! If your data set isn't very large, you probably don't need any aggregate tables.
4. Decide which aggregate tables to create. If you turn on SQL tracing, looking at the GROUP BY clauses of the long-running SQL statements will be a big clue here.
5. Register the aggregate tables in your catalog, create the tables in the database, populate the tables, and add indexes.
6. Restart Mondrian, to flush the cache and re-read the schema, then go to step 2 to see if things have improved.

AggGen 

AggGen is a tool that generates SQL to support the creation and maintenance of aggregate tables, and would give a template for the creation of materialized views for databases that support those. Given an MDX query, the generated create/insert SQL is optimal for the given query. The generated SQL covers both the "lost" and "collapsed" dimensions. For usage, see the documentation for CmdRunner.

Optimizing Calculations with the Expression Cache 

Mondrian may have performance issues if your schema makes intensive use of calculations. Mondrian executes calculations very efficiently, so usually the time spent calculating expressions is insignificant compared to the time spent executing SQL, but if you have many layers of calculated members and sets, in particular set-oriented constructs like the Aggregate function, it is possible that many thousands of calculations will be required for each cell.

To see whether calculations are causing your performance problem, turn on SQL tracing and measure what proportion of the time is spent executing SQL. If SQL is less than 50% of the time, it is possible that excessive calculations are responsible for the rest. (If the result set is very large, and if you are using JPivot or XML/A, the cost of generating HTML or XML is also worth investigating.)

It caches cell values retrieved from the database, but it does not generally cache the results of calculations. (The sole case where mondrian caches expression results automatically is for the second argument of the Rank(<Member>, <Set>[, <Expression>]) function, since this function is typically evaluated many times for different members over the same set.)

Since calculations are very efficient, this is generally the best policy: it is better for mondrian to use the available memory to cache values retrieved from the database, which are much slower to re-create.

The expression cache only caches expression results for the duration of a single statement. The results are not available for other statements. The expression cache also takes into account the evaluation context, and the known dependencies of particular functions and operators. For example, the expression

Filter([Store].[City].Members, ([Store].CurrentMember.Parent, [Time].[1997].[Q1])) > 100)

depends on all dimensions besides [Store] and [Time], because the expression overrides the value of the [Store] and [Time] dimensions inherited from the context, but the implicit evaluation of a cell pulls in all other dimensions. If the expression result has been cached for the contexts ([Store].[USA], [Time].[1997].[Q2], [Gender].[M]), the cache knows that it will return the same value for ([Store].[USA].[CA], [Time].[1997].[Q3], [Gender].[M]); however, ([Store].[USA], [Time].[1997].[Q2], [Gender].[F]) will require a new cache value, because the dependent dimension [Gender] has a different value.

However, if your application is very calculation intensive, you can use the Cache(<Expression>) function to tell mondrian to store the results of the expression in the expression cache. The first time this function is called, it evaluates its argument and stores it in the expression cache; subsequent calls within the an equivalent context will retrieve the value from the cache. We recommend that you use this function sparingly. If you have cached a frequently evaluated expression, then it will not be necessary to cache sub-expressions or super-expressions; the sub-expressions will be evaluated less frequently, and the super-expressions will evaluate more quickly because their expensive argument has been cached.

Author: Sherman Wood & Julian Hyde; last updated November, 2007.
Version: $Id$ (log)
Copyright (C) 2005-2007 Pentaho and others