Getpage-workload based index optimizer

[0001] This application claims priority of pending U.S. Provisional Patent Application Ser. No. 60/360,911, filed Mar. 1, 2002.

[0002] The present invention is an index optimization (XOP) for IBM mainframe DB2 database servers and the like. More specifically, the invention relates to an index recommendation program that is based on relevant SQL statements and their getpage activity as monitored by a DB2 trace for a representative period of time.

[0003] A DB2 database on an IBM mainframe server is a relational database management system. In a relational database, data is stored in one or more tables, each containing a specific number of rows. The rows of a table are divided into columns. Rows can be retrieved and manipulated at the column level. The language for DB2 data retrieval and manipulation is SQL (Structured Query Language). Instructions to the database for either retrieval or manipulation are hereinafter referred to as SQL statements or more generally, as queries. The major SQL operations commands are SELECT, INSERT, UPDATE, and DELETE. The performance of SQL statements can be enhanced through the use of indexes in accessing the rows in the DB2 tables more efficiently. An index is an ordered set of pointers to the data in the table i.e., it orders the values of columns in a table. Indexes are stored separately from the data. The decision on whether and how to use an existing index for a given SQL statement that retrieves/manipulates table data is made internally by the IBM DB2 optimizer. This decision is called the access path of an SQL statement. The access path and the performance of an SQL statement are directly related to index design. In large-scale production environments, thousands of tables may exist. Normally, index design is a manual process that includes estimates rather than actual production activity. The manual process of index design in such an environment is both inefficient and slow.

[0004] An SQL statement is internally subdivided into one or more query blocks by the IBM DB2 optimizer. Every query block has its own access path with one or more tables and may use one or more indexes, e.g., the query SELECT*FROM TAB1 UNION SELECT*FROM TAB2 consists of 2 query blocks SELECT*FROM TAB1 and SELECT*FROM TAB2. Other examples of query blocks are VIEW or SUBSELECT MATERIALIZATION. The query blocks and their selected access paths of SQL statements can be analyzed using the EXPLAIN interface of DB2.

[0005] Statistical data about the tables and indexes are a major influence on the access path of an SQL statement selected using the IBM DB2 optimizer. The DB2 catalog is an internal set of tables of data about a DB2 database including information on tables, columns, indexes, key columns, and SQL statements.

[0006] Buffer pools, also known as virtual buffer pools, are areas of virtual storage in which DB2 temporarily stores pages of data from tables and indexes. When an SQL statement requests rows of a table, the DB2 data manager retrieves the pages containing the rows of the table from the DB2 buffer pool manager. The number of pages retrieved for a table or index to satisfy an SQL statement is the “getpages” for the table or index.

[0007] The getpages measured in the buffers of the DB2 database server are always the same for an SQL statement and a given set of indexes on the tables referenced by the SQL statement, regardless of the state of the DB2 database server and the machine. In this regard, getpages are different from disk I/Os which may vary for the same SQL statement at different times.

[0008] Predicates are located in the WHERE, HAVING and ON clauses of SQL statements and describe attributes of table data. To reduce the table getpage activity, the IBM DB2 optimizer internally uses predicate filter factors to determine which index is best. Each such index can serve one or more columns of a table. A filter factor is a floating value between 0 and 1 that describes the proportion of rows in a table for which the predicate is true. It is supplied by the IBM DB2 optimizer EXPLAIN interface and stays the same regardless of the indexes deployed, i.e., filter factors are static for the predicates of a given query. A predicate with a small filter factor, i.e., close to zero is very effective because it selects only a few rows out of all rows in a table. A predicate with a high filter factor, i.e., close to 1, is very ineffective since it encompasses most rows in the table.

[0009] The following list shows the input data to the present invention. The abbreviations on the left hand side will be used herein:

• [0010] Q SQL statement
• [0011] T Table
• [0012] B Query block
• [0013] P Predicate
• [0014] FF Filter factor of a predicate
• [0015] Getp(T,Q) Number of getpages (equates to number of pages read) for table T for all invocations of SQL statement Q (regardless whether Q is a SELECT, UPDATE, INSERT or DELETE statement)
• [0016] Getp(I,Q) Number of getpages (equates to number of pages read) for index I for all invocations of SQL statement Q (regardless whether Q is a SELECT, UPDATE, INSERT or DELETE statement)
• [0017] Update(T,Q) Number of pages changed in table T by all invocations of an UPDATE statement Q
• [0018] Insert(T,Q) Number of pages changed in table T by all invocations of an INSERT statement Q
• [0019] Delete(T,Q) Number of pages changed in table T by all invocations of a DELETE statement Q
• [0020] Col(T,Q,B)={C₁, . . . , C_n} is the set of columns within table T and SQL statement Q and query block B
• [0021] Pred(T,Q,B)={P₁, . . . , P_n} is the set of predicates in SQL statement Q and query block B with a column in table T

[0022] All indexing n, such as P_n, should be understood as a generic placeholder that differs for each table, index and query block.

[0023] The present invention describes an index recommendation XOP program that computes new and improved performance indexes based on the getpage workload of a given set of SQL statements. The term getpage workload of an SQL statement is defined as the read/change activity (measured in getpages) on each table and each index currently used for all invocations of an SQL statement. The results of the XOP program are pure computations; the XOP program itself does not use the DB2 optimizer to determine the usefulness of potential indexes.

[0037] FIGS. 1 & 2 form a high level flowchart of the XOP program. For a given set of SQL statements, the programs major objective is to recommend new or improved performance indexes based on their ability to minimize the overall SQL statement getpage workload i.e., the read/change activity on each table and each index currently used for all invocations of an SQL statement.

[0038] The initial step to running the XOP program is to extract production statistics from the existing database. A conventional DB2 trace is also run in the production environment 10. This captures SQL statements and their getpage activities on the tables and index levels for a representative time interval. The trace data is expanded with static SQL from the database's catalog 11 and condensed to give the SQL frequency and to allow assignment of getpage counts to the SQL 12.

[0039] The test environment consist of corresponding tables and indexes to those in the production environment, but no data is required to be in the tables. The trace data is reloaded into the test environment 13. Production catalog statistics must be simulated in order to do a one time EXPLAIN of the relevant SQL statements. The EXPLAIN function of DB2 is used to provide the access path for each query block, the predicate filter factors, and the tables and indexes used by the relevant SQL statements. As a direct result, a comparison of indexes that exist with the ones being used, allows identification of the indexes currently not used 15. Next the program identifies the indexable predicates in each SQL, it then assigns a getpage count to each table in the SQL 14.

[0040] The actual selection of indexes to be used in the production database is done through a series of steps 200-800 which will determine the XOP selection process 16. As more fully discussed below, the tables are ranked 200, filter factors are obtained for all predicates 300, concatenation for all predicates as obtained 400, column permutations and filter factors and savings are calculated 500, column permutations are consolidated and indexes recommended 600, indexes for index only access are selected 800. Upon completion the recommended indexes are used to create indexes in the production database 17.

[0041] FIG. 3 illustrates the process of ranking tables. This ranking identifies the tables most likely to benefit as a result of new performance indexes. A set of tables TAB is introduced in the first step and initialized to empty 201. Then each table T 202 is checked whether its read activity Getp_T exceeds the change activity Chng_T by a predefined threshold percentage 203. A table T with a high volume of changed data is not considered for further processing because additional indexes on table T increase the change activity Chng_T to an amount that is higher than the savings in read activity Getp_T. Otherwise, table T is added to the set of tables TAB 204 which collects all tables for ranking.

[0042] The read activity Getp_T of a table T is defined as the total sum of getpage requests of all SQL statement invocations for this table T: ${\mathrm{Getp}}_T=\sum _{i=1}^N\mathrm{Gept}\left(T,Q_i\right)$

[0043] The change activity Chng_T of a table T is the total sum of pages changed of all UPDATE, DELETE and INSERT statement invocations for this table T: ${\mathrm{Chng}}_T=\sum _{i=1}^n\mathrm{Update}\left(T,Q_i\right)+\sum _{i=1}^n\mathrm{Delete}\left(T,Q_i\right)+\sum _{i=1}^n\mathrm{Insert}\left(T,Q_i\right)$

[0044] As one example, let a first SQL statement be SELECT*FROM TABA WHERE COLA=“A” with a read activity of 100 pages and let a second SQL statement be INSERT INTO TABA VALUES (“A”, “B”, “C”) with a read activity of 10 pages and a change activity of 100 pages. Let an additional one-column index using COLA reduce the read activity by half and double the change activity, then read activity is reduced to 50+5=55 pages and change activity is increased to 200 pages. The new overall workload is 55+200=255 pages which is an increase of 45 pages compared to the old workload of 100+10+100=210 pages.

[0045] Finally, all tables T in TAB are sorted by read activity Getp_T in descending order 205. The result is a table ranking that lists the tables having a high potential for workload savings from new performance indexes.

[0046] The flowchart of FIG. 4 shows the process of retrieving filter factors of all predicates. For each table T being processed in ranking order 301, for each SQL statement Q using this table T 302 and for each query block B inside this SQL statement Q 303, each predicate out of Pred(T,Q,B)={P_T1, . . . , P_Tn} 304 is analyzed. If the predicate is not indexable 305, its filter factor is set to 1 which prevents this predicate from having any influence on future index recommendations 307. Predicates that do not have any table columns will receive a filter factor of 1. Examples for such predicates are special registers or constants.

[0047] The flowchart of FIG. 5 illustrates the retrieval of the predicate concatenations, i.e., the boolean operators AND and OR. For each table T selected in ranking order 401, for each SQL statement Q using this table T 402, and for each query block B inside this SQL statement Q 403, each predicate out of Pred(T,Q,B)={P₁, . . . , P_n} 404 is investigated to determine how it is concatenated to the next predicate and whether the next predicate in the concatenation of SQL statement Q belongs to Pred(T,Q,B), i.e., contains columns of table T 405. If no such columns are found, such predicates are replaced by TRUE and removed from the predicate expression using boolean rules. Note that expressions in inner brackets are evaluated first.

[0048] As an example, let a predicate expression be A=1 AND B=5 AND C=1 OR D=1 where A, B and D are columns of a first table and C is a column of a second table. Then the expression for predicate set {A=1, B=5, D=1} can be rewritten as A=1 AND B=5 AND TRUE OR D=1 which is equivalent to A=1 AND B=5 OR D=1.

[0049] The next step of the program is shown in the flow diagram of FIG. 6. The objective is to build column sets that are input to the calculation of the total savings in getpages at table level for each column set permutation in a later step.

[0050] First, for each table T being processed in ranking order 501 a set of column sets ColComb_T is introduced in a first step and initialized as being empty 502. Then, for each SQL statement Q using this table T 503 and for each query block B inside this SQL statement Q 504, the set of columns Col(T,Q,B)={C₁, . . . , C_n} is checked to determine whether column set {C₁, . . . , C_n} is already an element in ColComb_T 506. If it is not, column set {C₁, . . . , C_n} is added to ColComb_T 507.

[0051] When all SQL statements and all their query blocks are processed for a given table T, i.e., the build up of ColComb_T has finished for given table T, column sets Col(T,Q,B) inside ColComb_T are pruned 508 as seen more fully in the flow diagram of FIG. 8. Column permutations are built and their filters and savings in getpages for each query block B are computed. The purpose of column set pruning is to avoid redundant combinations in the column set permutations that are done in a later step 711. For each column set {C₁, . . . , C_n} out of ColComb_T 701 it is checked whether it is a real subset of another column set {d₁, . . . , d_x} out of ColComb_T 702. If this is true, the element {C₁, . . . , C_n} is removed from ColComb_T 703.

[0052] As an example, let {A,B,C}, {A,C} and {D,E} be the elements in ColComb_T, the element {A,C} will be removed from ColComb_T because it is a subset of {A,B,C}.

[0053] As seen in the flow diagram of FIG. 6, the program then verifies whether there is enough read activity Getp_T left over for all tables T not yet processed. If the remaining read activity is sufficient according to a predefined threshold, the process of building column sets and the calculation of the total savings in getpages on table level for each column set permutation continues for the next table T 501.

[0054] The savings Sav_T,Q,B,Pe for each permutation Pe and query block B inside every SQL statement Q using table T are known 704. Thus, the total saving Sav_T,Pe for table T and each permutation Pe (705) can be computed 706 as: ${\mathrm{Sav}}_{T,\mathrm{Pe}}=\sum _{i=1}^n\sum _{j=1}^n{\mathrm{Sav}}_{T,Q_i,B_j,\mathrm{Pe}}$

[0055] The program verifies new indexes are recommended based on the calculated savings Sav_T,Pe for all processed tables T as shown in the flow diagram of FIG. 7. First, the savings Sav_T,Pe for all permutations Pe and all tables T are sorted in descending order 601, so that the permutations Pe with the highest savings in getpages are first. For each saving Sav_T,Pe being processed in sort order 602, it is verified whether it exceeds a predefined threshold value 603. If so, a recommendation is made for a new index with a key Pe=(E₁, . . . , E_n). Depending on the getpage activity caused by ORDER BY, MIN and MAX the index key columns E₁, . . . , E_n are put in either ascending or descending order. To avoid redundant index recommendations, the sorted list is searched for elements Sav_T,Px of the same table T where the permutation Px is covered by the index key (E₁, . . . , E_n) just recommended. Such elements Sav_T,Px are ignored for further processing and removed from the sorted list 602.

[0056] As an example, let (A, B, C) be the key columns of the recommended new index, permutations (A) and (A, B) do not need any further attention and are deleted because they are already covered by index key (A, B, C).

[0057] The flowchart 508 of FIG. 8 shows returning the savings Sav_T,Q,B,Pe for each permutation Pe and query block B inside every SQL statement Q using table T 704 FIG. 7:

[0058] First, for each column set Col(T,Q,B)={C₁, . . . , C_n} inside ColComb_T 710 all column permutations are built 711. As one example, let {A, B, C} be the column set. Then there are 15 column permutations which are (A), (B), (C), (A, B), (A, C), (B, A), (B, C), (C, A), (C, B), (A, B, C), (A, C, B), (B, A, C), (B, C, A), (C, A, B) and (C, B, A).

[0059] Then, for each permutation Pe 712, all query blocks B with a column in Table T for any SQL statement Q 713 are processed in the following manner:

• [0060] 1. The filter factor FF_T,Q,B,Pe for query block B and permutation Pe and the filter factor FF_T,Q,B for the complete query block B are calculated 714.
• [0061] 2. The getpage number Getp_T,Q,B on query block level B is calculated based on FF_T,Q,B 715.
• [0062] 3. Depending on the current index usage 716, 717 of query block B, the savings Sav_T,Q,B,Pe for permutation Pe is computed 718, 719, 720.

[0063] For each predicate set Pred(T,Q,B) and its concatenations (modified as described in FIG. 4), the overall filter factor FF_T,Q,B is calculated using the following boolean AND and OR rule:

• • [0064] Filter factors of predicates that are concatenated with an AND will be multiplied.
  • [0065] Filter factors of predicates that are concatenated with an OR will be added and the product of both filter factors is subtracted.
  • [0066] Filter factors of predicate expressions in inner brackets are computed first.
    The above rule will be referenced in the following description as Rule_ANDandOR. As a continuation of the example already given, let a predicate expression be A=1 AND B=5 OR D=1 where A, B and D are all columns of table T. Let FFa, FFb and FFd be the filter factors of the columns A, B and D. Then FF_T,Q,B is computed as ((FFa*FFb)+FFd)−((FFa*FFb)*FFd). In the next step, only those predicates and concatenations are kept in the predicate expression that contain a column in permutation Pe. All others are set to TRUE and removed using boolean rules. The filter factor FF_T,Q,B,Pe is now calculated using Rule_ANDandOR. Note that expressions in inner brackets are evaluated first. In the above example, let (A, D) be the permutation Pe. Then predicate expression A=1 AND B=5 OR D=1 equates to an expression A=1 OR D=1 and filter factor FF_T,Q,B,Pe is computed as FFa+FFd−(FFa*FFd).

[0067] The number of getpages Getp_T,Q,B for table T and query block B in SQL statement Q is computed as: ${\mathrm{Getp}}_{T,Q,B}=\mathrm{Getp}\left(T,Q\right)*\frac{{\mathrm{FF}}_{T,Q,B}}{\sum _{i=1}^n{\mathrm{FF}}_{T,Q,\mathrm{Bi}}}$
where Bi are the query blocks in SQL statement Q.

[0068] The savings Sav_T,Q,B,Pe for permutation Pe is calculated in one of the following three ways:

[0069] Formula 1:

[0070] The query block B currently uses a matching index access 718: ${\mathrm{Sav}}_{T,Q,B,\mathrm{Pe}}=\frac{{\mathrm{Getp}}_{T,Q,B}}{\mathrm{FilterFactorI}}*\left(1-{\mathrm{FF}}_{T,Q,B,\mathrm{Pe}}\right)$
where FilterFactorI stands for the filter factor of the used index. It is calculated for a specific query block B with the same AND and OR rule already described. Assume Pred_i are the predicates in query block B with columns covered by the index, then the FilterFactorI is calculated as follows:

• • [0071] FilterFactorI=Rule_ANDandOR[FF(Pred₁) Concat(Pred₂) FF(Pred₂) . . . FF(Pred_n−1) Concat(Pred_n) FF(Pred_n)]
    where FF(Pred_i) are the filter factors of Pred_i and Concat(Pred_i+1) are the concatenations AND and OR to the next predicate Pred_i+1.

[0072] Formula 2:

[0073] The query block B currently uses a non-matching index access 719:

• • [0074] Sav_T,Q,B,Pe=(Getp_T,Q,B+Getp(I,Q))*(1−FF_T,Q,B,Pe)
    where Getp(I,Q) is the number of index getpages. Normally, Getp(I,Q) exceeds Getp_T,Q,B by far and will be theoretically eliminated by a new index with matching access.

[0075] Formula 3:

[0076] The query block B does not currently use any index 716, i.e., uses a table space scan 720:

• • [0077] Sav_T,Q,B,Pe=Getp_T,Q,B*(1−FF_T,Q,B,Pe)

[0078] All permutations Pe 712 are processed for all column sets Col(T,Q,B)={C₁, . . . , C_n} inside ColComb_T 710. Now, the total saving Sav_T,Pe for table T and permutation Pe can be computed as the sum of all Sav_T,Q,B,Pe 706.

[0079] The final steps in the XOP program is shown in flowchart of FIGS. 10-12 where recommendations are made for new or extended indexes which allow index only access results for selected SQL statements. Index only access means that SQL statements are solely processed by retrieving index pages. No pages have to be read on the table level because all SQL statement columns are part of the index.

[0080] First, each table T being processed in ranking order 801 is checked whether its read activity Getp_T exceeds the total read activity of all tables by a predefined threshold percentage 802. If this threshold is not reached, this table T and all tables that follow table T in ranking order are not considered any further and processing continues with flowchart FIG. 13809. Otherwise, if the read activity on table T is still sufficient, there should be no index recommendation already made for this table T by the XOP program 803. If this is true, all existing indexes I for table T are retrieved (804) and sets AddCol_T,I are initialized to “empty” for table T and all indexes I.

[0081] A set AddCol_T,I will store elements (B,{C₁, . . . , C_n},Getp_T,Q,B) which represent the column combination {C₁, . . . , C_n} and its savings Getp_T,Q,B in getpages for a query block B 805 inside SQL statements Q using table T (806) that is to be added to index I for index only access. The construction of AddCol_T,I is illustrated in flowchart FIG. 11 where all indexes 810 are investigated for query block B which columns {C₁, . . . , C_n} are to be added 811 to the index key for index only access. If the total length and the total number of the index key and additional columns do not exceed DB2 limits and if DB2 supports index only access for any additional column variable in length 812, then the column combination {C₁, . . . , C_n} is stored together with its savings in set AddCol_T,I 813 for index I and table T, i.e., AddCol_T,I=AddCol_T,I ∪{(B,{C₁, . . . , C_n},Getp_T,Q,B)}. When all indexes I are checked for query block B (810), processing returns to flowchart FIG. 8807.

[0082] As an example, let the columns of table T be A, B, C, D, E, F, G, and H. Let the 3 indexes on this table be:

[0083]

	Index:	I1	I2	I3
	Key:	(B, C)	(D, G)	(E)

[0084] Furthermore, let the columns of 6 query blocks with columns in T be COL(T,Q,B₁)={B, C, D}, COL(T,Q,B₂)={B, D, G}, COL(T,Q,B₃)={B, C, F}, COL(T,Q,B₄)={B, C, D}, COL(T,Q,B₅)={B, C, D, E, F}, and COL(T,Q,B₆)={F}. Then, the sets AddCol_T,I1, AddCol_T,I2, and AddCol_T,I3 will contain the following elements:

[0085]

	AddColT,I1	AddColT,I2	AddColT,I3
{B, C, D}:	(B1, {D}, 30)
{B, D, G}:		(B2, {B}, 70)
{B, C, F}:	(B3, {F}, 10)
{B, C, D}:	(B4, {D}, 20)
{B, C, D, E,	(B5, {D, E, F},		(B5, {B, C, D, F},
F}:	30)		10)
{F}:

[0086] When the sets AddCol_T,I are built for a given table T and all its indexes I 805, processing continues with flowchart FIG. 12808 where the column combination found before is pruned. Here, for each index I on table T 820, each element (B,{C₁, . . . , C_n},Getp) inside AddCol_T,I is checked 821 whether {C₁, . . . , C_n} covers any other element inside AddCol_T,I 823, i.e., {d₁, . . . , d_n} is a subset of {C₁, . . . , C_n} and (B_d,{d₁, . . . , d_n},Getp_d)∈AddCol_T,I. If this is true, the read activity Getp_d is added to the overall savings Getp_T,I,B that an index I on table T will have if columns {C₁, . . . , C_n} are added to its index key 824. The overall savings Getp_T,I,B is initially set to Getp 822.

[0087] For a given table T and index I, the element (B,{C₁, . . . , C_n},Getp) with the maximum Getp_T,I,B finally remains in AddCol_T,I 825 and all AddCol_T,I elements for other query blocks B_d covered by {C₁, . . . , C_n} are removed from AddCol_T,I. Furthermore, all elements for query blocks B and B_d are removed from the sets AddCol_T,Ix for any other Index Ix 826.

[0088] Processing returns to flowchart FIG. 10 at 808 if all indexes I on table T are checked 820 or the sum of getpages in the remaining query blocks is below a predefined threshold 827.

[0089] When all tables are processed 801 or the read activity Getp_T of the next table T does not exceed the total read activity of all tables by a predefined threshold percentage 802, processing continues with flowchart FIG. 83 809. Here the indexes are recommended that support index only access. For every element (B,{C₁, . . . , C_n},Getp) out of AddCol_T,I and its overall savings Getp_T,I,B 830, it is checked whether Getp_T,I,B exceeds the read activity Getp_T of table T by predefined threshold 831. If the savings Getp_T,I,B are not sufficient, the next element (B,{C₁, . . . , C_n},Getp) is checked 830. Otherwise, it is decided whether a new index should be built with a key containing the columns copied from the existing index I plus the columns {C₁, . . . , C_n} 833 or whether the columns {C₁, . . . , C_n} should be just added to the existing index I 834. To avoid changes to the data model, the recommendation for a new index is always made if the existing index is unique or clustering 832. Otherwise, recommendations are made to modify the existing index 834.

[0090] The recommendations can then be used to create a indexes in the production database 17 which are accordingly optimized based on getpage work load.

[0091] It is understood that the present embodiments described above are to be considered as illustrative and not restrictive. It will be obvious to those skilled in the art to make various changes, alterations and modifications to the invention described herein. To the extent that these variations, modifications and alterations depart from the scope and spirit of the appended claims, they are intended to be encompassed therein.