Understanding DB2 Optimizer

Editor's Notes

• #18: What are we trying to Explain? We want to explain how our query goes about fulfilling our request. Why is this important? Because there are many ways of going about fulfilling the request. Some are faster or more efficient than the others (in terms or resource usage). This about a GPS device… Its sole purpose is to get you from point A to point B in the fastest possible time. The optimizer is the same. This “way” of fulfilling the request is called an access plan. Namely, an access plan specifies a type and order of operations for accessing the data, and producing a result set. The optimizer produces an access plan whenever you compile a SQL statement. We also refer to access plan also as: explain = db2exfmt = access path What is the purpose of the Explain tools? The explain tools produce visual representation of access plan, and they dump certain information that the access plan is based upon. Explain output lets you view statistics for selected tables, indexes, or columns; properties for operators; global information such as table space and function statistics; and configuration parameters relevant to optimization. It lets you evaluate your query performance tuning options. Perform "what-if" query performance analysis. Once you become familiar with the explain tool, it will become your first choice to evaluate a query performance problem.
• #19: Sequence of operations to process the query: How did we get from point A to point B. What operations were done from the time the query was submitted to the time the result set is returned. Cost information: Everything that the access plan does with data along the way has certain cost in terms of CPU cycles, input and output performed on memory or disk, or both. This cost is expressed in timerons. Predicates and selectivity estimates for each predicate: These are the conditions that we used to limit the search scope of our query, along with estimates of how much data the condition will filter through. Statistics for all objects referenced in the SQL query at the time the explain information is captured As we shall see, statistics are critical for the optimizer decision making process. Hence they are contained in the explain output.
• #20: The Explain tables capture access plans when the Explain facility is activated. This is where the access plain information goes for the explain purposes. Not to a file or memory.
• #21: Tables are listed here for your reference.
• #22: Db2expln: Provides basic, relatively compact and English-like overview of access plan and the related data. Unlike db2exfmt, does not make recommendations on what to do to improve performance. Use when you want a quick and dirty explanation of what kind access plan was used (such as table access and join information). Db2exfmt: Provides more detailed information, including the catalog statistics (such as column distribution stats) which are crucial to the optimizer’s decision making process. We will be using db2exfmt as the primary explain tool in this workshop. Recommendation: strive to use it always over db2expln. Since you are going through the trouble of generating explain output, you might as well run db2exfmt for the sake of completeness.
• #23: These are the sections of the db2exfmt output. We will briefly discuss each section.
• #24: Explain instance: Contains general information about the environment, such as: DB2 version, time, schema name, etc. Database context: List instance and database configuration parameters that were in effect during the query compilation. These parameters are important because they influence the access plan generation. Package context: List query type (static or dynamic ) and optimization level. The optimizer level is important because it determines how hard the optimizer will work to produce an optimal access plan. Original statement: Actual query text. Optimized statement: Rewritten query text according to the applied optimization techniques.
• #25: Access plan: This is the big one… May want to have an full HD monitor. Graphical representation of the query execution. Displays the order of operations along with rows returned, cost and I/O for each operation. Execution starts at the bottom, and it goes up. Next, we are onto the advanced sections. Extended diagnostic information: Warnings, informational or error messages returned by the optimizer. Very important section as it may warn you about missing stats. For example, this section may tell you that your query may perform poorly because you have not collected distribution stats on a column B. Plan details: Detailed information for each operator including arguments, predicates, estimates (filter factor), and input/output streams. This section tells us what did the optimizer based its calculations on. Objects used in the access plan: Any objects (table/index) used by the query, and/or the access plan is displayed on this are which includes statistical information.
• #26: Syntax diagram for db2exfmt for the participants reference.
• #27: The most important options for our purposes: NO Disables the Explain facility. No Explain information is captured. NO is the initial value of the special register. YES Enables the Explain facility and causes Explain information to be inserted into the Explain tables for eligible dynamic SQL SQL query. All dynamic SQL SQL query are compiled and executed normally. EXPLAIN Enables the Explain facility and causes Explain information to be captured for any eligible dynamic SQL query that is prepared. However, dynamic SQL query are not executed.
• #31: An operator typically takes one or more inputs, does something with it, and passes the output to the next operator.
• #32: An operator typically takes one or two input streams. In this example, we see three operators: two table scans and a join. Two TBSCAN operators each perform a table scan and feed the results into the third operator – the HSJOIN. The third operator combines the two data streams called using the join type called hash join, and it passes the resulting data stream into the next operator, and so on. Remember that the figures (rows returned and cost) are estimates.
• #33: TBSCAN and IXSCAN will typically be placed at the bottom of the access plan graph. These are the operations that retrieve data from the sources (tables and indexes). Recall that graphical representation of execution starts at the bottom, and it goes up.
• #34: Some of the common operators are the join operators: HSJOIN, MSJOIN and NLJOIN. These operators merge two streams of data. At least one of them will be found in any typical query which queries multiple tables.
• #35: SORT is another common operator Sorting is required when no index exists that satisfies the requested ordering, or when sorting would be less expensive than an index scan. Sorting is usually performed as a final operation once the required rows are fetched, or to sort data prior to a join or a group by.
• #36: Another common operator, especially in more complex queries: TEMP The action of storing data in a temporary table, to be read back out by another operator (possibly multiple times). This operator is typically required to evaluate subqueries or to store intermediate results.
• #39: We have learned what the optimizer is. Now, we are getting into details about how it works.
• #40: Why is the optimizer heavily influenced by the statistical information? Because the optimizer uses this information to estimate the costs of alternative access plans for each query and choose the one. As we shall see, there are other factors that influence the optimizer, but the statistics form the basis for access plan calculations. Going back to the GPS example… To compare the alternatives, a GPS device needs to know the length of each road, their speed limits, and current traffic conditions. In DB2 context we refer to these figures as statistics. Statistics describe characteristics of database objects (e.g., tables and indexes). Distribution statistics are used to estimate how many rows will be returned. This is important because the optimizer chooses join type or sort method based on the distribution stats.
• #41: Examples of statistical information used by the optimizer: table and index cardinality the distribution of data in specific table columns the cluster ratio of indexes the number of leaf pages in indexes the number of table rows that overflow their original pages the number of filled and empty pages in a table and so forth.
• #42: In the nutshell, the RUNSTATS utility collects statistics about tables, indexes and statistical views, and stores them into catalog tables for use by the optimizer. We want to run RUNSTATS after a significant change to data (either volume or cardinality) because the stats get out of date. The main risk of having outdated stats is that the compiler may select a suboptimal plan for a given query. For instance, the optimizer may estimate based on stale statistics, that a query would return 100 rows from table T1. As a result it chooses nested loop join to join the tables. However, in reality the query returns one million rows from T1, for which merge join may have been more suited. As a result, the query that used to run fast now grinds to a halt. Statistical view: special type of view which combines the columns referenced by commonly used queries, and which enabled for use in query optimization.
• #43: NUM_FREQVALUES Defines the maximum number of frequency values to collect. NUM_QUANTILES Defines the maximum number of distribution quantile values to collect. You might also want to collect statistics on system catalog tables, if an application queries these tables directly and if there is significant catalog update activity, such as that resulting from the execution of data definition language (DDL) statements.
• #44: Because these operations add significant amount new data (such as LOAD and IUD), or they change data layout on disk (REORG, REDISTRIBUTE), or the way data is accessed (INDEX creation).
• #46: Distribution statistics make the optimizer aware of data skew. Namely, how many times a given value appears in the column. If distribution stats are so good for the optimizer, why not collect them all the time for every column, and do it five times a week? Because RUNSTATS is not free. It consumes system resources and contends with queries for CPU cycles and buffer pools. It also places locks on database objects, which may reduce concurrency (although RUNSTATS can be throttled to have low priority and yield to applications when there are locking conflicts). Detailed index statistics provide more details about the I/O required to fetch data pages when the table is accessed using a particular index. However, collecting detailed index statistics consumes considerable processing time and memory for large tables. As a result, you should schedule RUNSTATS execution during maintenance windows, typically after REORG.
• #64: There are also joins based on a cluster of operators (i.e. starjoin, order NLJOIN) The are joins specific to DPF (i.e. collocated joins, directed joins, broadcast joins, repartitioned joins).
• #85: Database manager configuration: These options affect all databases in a given instance. May require instance restart after setting a new parameter value. Database configuration: These options affect only the database for which they are defined. May require database restart after setting a new parameter value. Query compiler registry variables: Affect or more instances on the system. Typically require instance restart. CURRENT QUERY OPTIMIZATION special register: Affect only the current session.
• #86: Parallelism INTRA_PARALLEL: Tells DB2 to take advantage of symmetric multiprocessing (SMP) on machines that have two or more CPUs or CPU cores. Not everything will run in parallel. It enables some parts of query execution to run in parallel within a single database partition. For example, index sort will run in parallel if you have this parameter enabled. Index keys will be fetched, then multiple sort agents would be started, one for each CPU/core. When all the sort passes are complete, these sorted subsets are merged into a single sorted set of index keys. Also, HSJOIN runs in parallel (i.e., build table partitioned in parallel and loaded into memory). Why is parallelism important to the optimizer? Because it may favour certain operators over the others when parallelism is on (i.e., HSJOIN over NLJOIN). CPU Speed (CPUSPEED): The optimizer uses the CPU speed to estimate the cost of performing certain operations. This configuration parameter is automatically set to an appropriate value when the database is installed or upgraded. Recommendation: do not adjust this parameter unless you are modelling a production environment on a test system or assessing the impact of a hardware change. Using this parameter to model a different hardware environment allows you to find out the access plans that might be chosen for that environment. To have the database manager recompute the value of this automatic configuration parameter, set it to -1. Communications speed (COMM_BANDWIDTH): The value calculated for the communications bandwidth, in megabytes per second, is used by the optimizer to estimate the cost of performing certain operations between the database partition servers of a partitioned database system. The optimizer does not model the cost of communications between a client and a server, so this parameter should reflect only the nominal bandwidth between the database partition servers, if any. Recommendation: You should only adjust this parameter if you wish to model a production environment on your test system or to assess the impact of upgrading hardware.
• #87: Buffer pools (BUFFPAGE): Represents the total size of the buffer pools that you specified when you created or altered them. Why is this parameter important? Because when the optimizer chooses the access plan, it considers the I/O cost of fetching pages from disk to the buffer pool and estimates the number of I/Os required to satisfy a query. The estimate includes a prediction of buffer-pool usage, because additional physical I/Os are not required to read rows in a page that is already in the buffer pool. The I/O costs of reading the tables can have an impact on: How two tables are joined Whether an unclustered index will be used to read the data What this tells is that bigger is better when it comes to the buffer pool size. However, there is a point of diminishing returns, so should be careful not to increase your buffer pools to 75% of your system memory because it may affect other processes.
• #88: Sort Heap Size (SORTHEAP): Defines the maximum number of memory pages to be used for sort operations. Why is this important? Because it helps the optimizer determine whether a sort will be performed in memory or be “spilled” to disk (i.e., to a system temporary table) due to running out of sort heap space. A sort that occurs in memory versus a system temporary table always results in better performance, and is used if possible. When choosing an access plan, the optimizer estimates the cost of the sort operations by: Estimating the amount of data to be sorted Looking at the sortheap parameter to determine if there is enough space to read a sort in a single, sequential access. Accurate costing of the sort operation helps the optimizer decide whether to use sort or index scan in certain cases. Database heap size (DBHEAP): There is one database heap per database, and the database manager uses it on behalf of all applications connected to the database. It contains control block information for tables, indexes, table spaces, and buffer pools. It is important to the optimizer because it helps estimate the amount of memory available to certain operations. Lock list size (LOCKLIST): Indicates the amount of storage allocated to the lock list. Maximum lock list (MAXLOCKS): Represents the maximum percentage lock memory that can be held by any one application before lock escalation is performed. Lock escalation, by the way, is the process of replacing a large number of row-level locks with a single table lock. It is done to reduce the amount of memory used to keep track of locks. Why are these parameters important? Because, when the isolation level is repeatable read (RR), the optimizer considers the values of the locklist and maxlocks parameters to determine whether row locks might be escalated to a table lock. If the optimizer estimates that lock escalation will occur for a table access, then it chooses a table lock for the access plan. This is done to prevent incurring the overhead of lock escalation during the query execution. Downside: while we avoid the overhead of escalation, table access will lock the entire table, thereby increasing the contention and possibly reducing database access concurrency. Recommendation: do some benchmarking to determine the minimum value for each parameter necessary to completely eliminate lock escalation.
• #89: Average applications (AVG_APPLS): The optimizer uses the avg_appls parameter to help estimate how much of the buffer pool might be available at run-time for the access plan chosen. Higher values for this parameter can influence the optimizer to choose access plans that are more conservative in buffer pool usage. If you specify a value of 1, the optimizer considers that the entire buffer pool will be available to the application. Optimization Level (DFT_QUERYOPT): Tells the optimizer which level of optimization to use by default when compiling queries (the default is 3). The higher the level, the more complex algorithms are considered. This means the optimizer will try more access plan combinations before choosing the best one. In general, a higher degree results in longer compile time. Query Degree (DFT_DEGREE): Specifies the degree of parallelism to be used by default within the single database partition . A value of one (1) means no intra-partition parallelism. A value of minus one (-1) or ANY means the optimizer determines the degree of intra-partition parallelism based on the number of processors and the type of query. Note: Intra-parallel processing does not occur unless you enable it by setting the intra_parallel database manager configuration parameter.
• #90: Number of frequent values retained (NUM_FREQVALUES): Allows you to specify the number of "most frequent values" that will be collected when the WITH DISTRIBUTION option is specified with the RUNSTATS command. The "most frequent value" statistics help the optimizer understand the distribution of data values within a column. A higher value results in more information being available to the query optimizer but requires additional catalog space and incurs higher heap space usage during RUNSTATS. When 0 is specified, no frequent-value statistics are retained, even if you request that distribution statistics be collected. Number of quantiles retained (NUM_QUANTILES): Controls the number of quantiles that will be collected when the WITH DISTRIBUTION option is specified on the RUNSTATS command. Each quantile represents to the subset of a distribution data containing one fifth (20%) of the total sample. The "quantile" statistics help the optimizer understand the distribution of data values within a column. A higher value results in more information being available to the query optimizer but requires additional catalog space. When 0 or 1 is specified, no quantile statistics are retained, even if you request that distribution statistics be collected.
• #91: DB2_EXTENDED_OPTIMIZATION: This variable specifies whether or not the query optimizer uses optimization extensions to improve query performance. It affects one of more instances on the system, along with all the databases within them. Use this if you wish to try specific optimization algorithms on your query (refer to the documentation for the description of algorithms). These optimization extensions may increase compilation time, and may not improve query performance in all situations. Testing should be done to determine individual query performance improvements. DB2_SORT_AFTER_TQ: Specifies how the optimizer works with directed table queues in a partitioned database environment when the receiving end requires the data to be sorted and the number of receiving nodes is equal to the number of sending nodes. When DB2_SORT_AFTER_TQ=NO, the optimizer tends to sort at the sending end and merge the rows at the receiving end. When DB2_SORT_AFTER_TQ=YES, the optimizer tends to transmit the rows unsorted, not merge at the receiving end, and sort the rows at the receiving end after receiving all the rows. DB2_INLIST_TO_NLJN: Sometimes the optimizer does not have accurate information to determine the best join method for the rewritten version of the query. This can occur if the IN list contains parameter markers or host variables which prevent the optimizer from using catalog statistics to determine the selectivity. This registry variable causes the optimizer to favor nested loop joins to join the list of values, using the table that contributes the IN list as the inner table in the join.
• #92: Use this statement when you want to test or apply a certain query optimization class to one or more queries in the current session, without affecting other database users. Default optimization class will be used for all other sessions.
• #99: What is optimization profile? Optimization profiles allow certain optimization guidelines to be provided to the optimizer without application or database configuration changes. Advanced method of query performance tuning. Why would we want to override the optimizer? After all, isn’t it smarter that we are in calculating hundreds of combinations in a split second and figuring out which one forms the best access plan? Yes and no… Going back to our GPS example: Imagine wanting to get from one end of Toronto to the other. There are countless possibilities. The shortest route is not always the fastest. Also, what is normally the fastest route may not be available to you that day due to road closure. In this case, you would want to override the GPS device and tell it to take a particular detour. The same logic applies to the optimizer. It may have the computational ability, but unlike a human brain it does not have a learning capability to make heuristic decisions on the fly. When you would want to use the optimizer profiles: You know your statistics are out of date, can’t afford to collect then now, but you want to tell the optimizer to use a particular index even though stats are telling us otherwise. You want to test various join combinations or table access methods (e.g., MQT, index) for your query. You want to apply different optimization techniques to a handful of queries without affecting all the other queries. You have exhausted all other tuning options.
• #100: XML document Why XML? XML is handy for describing data, so it was a logical choice for implementing optimizer profiles. It must conform to a certain XML schema, otherwise the optimizer returns a warning at compile time and it ignores the profile. Profile Header: Typically specifies schema version. Does not change a lot between the released, but you should always use the releases that corresponds to your DB2 level (e.g., V9.7.0.0). Global optimization guidelines: This section contains optimization guidelines that are to be used for all statements for which the profile is currently in effect. Statement optimization guidelines This section contains optimization guidelines that are to be used for a particular SQL statement only, e.g.: “For app1.0, only consider routing to MQTs: Newt.AvgSales and Newt.SumSales” “Use index ISUPPKEY to access SUPPLIERS in the subquery of query 9”
• #101: Optimization guidelines do not need to be comprehensive, but they should be targeted to a desired set of queries and a particular execution plan. The DB2 optimizer still considers other possible access plans using the existing cost-based methods. In this example: In the global guidelines section we are telling the optimizer to first consider MQT tables for all queries against the base tables Test.AvgSales and Test.SumSales for which the profile is currently in effect. In the statement guidelines section we are telling the optimizer to access table TAB1 through the I_SUPPKEY index, but only for the query specified in the STMTKEY tag.
• #102: More details on the statement section as it is here where you would place most of our guidelines. Each statement profile contains a statement key (STMTKEY) and one or more statement-level optimization guidelines (OPTGUIDELINES). The original statement and statement in the profile must be exactly the same, case sensitive. Otherwise, the optimizer ignores the profile. You may use the EXPLAIN statement to verify whether the optimizer applied the guideline (more on this later).
• #103: These are some examples of guidelines. There are no specific recommendations as to when you should use these guidelines. It depends on case-by-case basis. It is often lent to a trial and error approach. You may refer to the DB2 documentation for a more detailed description of query rewrite guidelines. http://publib.boulder.ibm.com/infocenter/db2luw/v9r7/topic/com.ibm.db2.luw.admin.perf.doc/doc/c0024581.html
• #104: First, you compose a simple XML document. Then, you insert it into the database. Finally, you specify the name of the optimization profile when you decide to use it. Limitation: The original statement and the statement in the profile must be exactly the same (case sensitive). Review the file prof1.xml
• #106: The optimizer ignores invalid or inapplicable optimization guidelines. If any optimization guidelines are ignored, an execution plan is created and SQL0437W with reason code 13 is returned. You can then use the EXPLAIN statement to get detailed diagnostic information regarding optimization guidelines processing. The Profile Information section contains the names of profile and guidelines in effect, as well as any warnings (e.g., SQL0437W)