Thursday, December 30, 2010

DB2 for z/OS: KEYCARD Gets its Due

By now, you've probably seen and/or heard a good deal of information about DB2 10 for z/OS, which was announced and became generally available this past October. There is indeed a lot of big news associated with this latest release of IBM's mainframe relational database management system: reduced CPU costs, support for temporal data (tables with system and/or business time dimensions), a huge increase in the number of threads that can be concurrently active, a migration path to universal tablespaces, SQL Procedure Language enhancements (better performance, plus the ability to write user-defined functions using SQLPL), table-based data access control policies, and lots more (and I do mean "lots" -- check out the DB2 10 "What's New" manual). Amongst all the "wow" features of DB2 10 are some nice little nuggets of functionality that will positively impact system performance and administration. One of these -- and the subject of this blog entry -- concerns the KEYCARD option of the RUNSTATS utility.

Here's the story in a nutshell: with DB2 10, the KEYCARD option is no longer optional. That is to say, if you run the RUNSTATS utility with an INDEX specification in a DB2 10 environment, and you have indexes with more than two key columns defined on the table (or tables) in the target tablespace, you will get the catalog statistics associated with KEYCARD, regardless of what you specify on the utility control statement. You can still put KEYCARD in this statement, but the keyword will be ignored because it "is now built into the normal execution of the RUNSTATS INDEX utility and cannot be disabled." Trust me: this is a good thing. It's a change that leaders on the optimizer team in IBM's DB2 development organization have lobbied for for quite some time. I'll explain why, and I'll tell you what this means for you if you're not yet running DB2 10 (and that's most of you, as, again, DB2 10 has only been generally available for a couple of months).

First, a little background. KEYCARD (which is only valid in the context of an INDEX specification on a RUNSTATS utility control statement) goes way back. I'm not sure when it was introduced as a RUNSTATS option, but I think that it might have been delivered with DB2 for z/OS Version 4 (mid-1990s). What does it do? Pretty simple: it causes RUNSTATS to determine the number of distinct values of combinations of an index's first "n" key columns, where "n" is greater than 1 and less than the total number of the index's key columns (I say this because the cardinality -- the number of distinct values -- of the index's first key column and of the full key are gathered anyway and placed in the FIRSTKEYCARDF and FULLKEYCARDF columns, respectively, of the SYSIBM.SYSINDEXES catalog table). The values obtained via the KEYCARD specification are placed in the SYSIBM.SYSCOLDIST table.

An example might be helpful here. Suppose you have a 10,000-row table containing data about customers in a particular country (one row per customer), and on that table you have an index with a key comprised of the columns STATE, CITY, and POSTAL_CODE. Suppose further that you have customers in fifty cities, with each city being in a different state (i.e., one city in each of 50 different states -- admittedly, this is a rather contrived example). Finally, assume that there are 200 different zip code values in the table, and that the duplicate values of STATE, CITY, and POSTAL_CODE are evenly spread across the table's rows (for information on non-uniform distribution of non-unique column values, check out an entry I posted last year on that topic in my old Catterall Consulting blog). Without KEYCARD specified, an execution of RUNSTATS will generate (as previously mentioned) FIRSTKEYCARDF (50) and FULLKEYCARDF (200) statistics for the index. With KEYCARD specified, RUNSTATS will also determine the number of distinct values of the index's first two columns: in this case, that's 50 -- the same as the FIRSTKEYCARDF value (if the index had 4 columns, KEYCARD would result in a determination of the number of distinct values of the combination of key columns 1, 2, and 3, as well as the number of distinct values of key columns 1 and 2).

Why is getting this statistical data important? Two words: column correlation. See, if you use what is a fairly common "base" RUNSTATS control statement, namely, TABLE(ALL) INDEX(ALL), DB2 will gather cardinality stats for every column of every table in the target tablespace. It will know, then, that there are 50 distinct values in the STATE column and 50 distinct values in the CITY column. What DB2 doesn't know is that the STATE and CITY values are highly correlated (CITY value 'Los Angeles' is paired with 'California', but not with any other value in the STATE column). Not knowing this, DB2 assumes that the values in the CITY column are independent of STATE values. Consider a query with the compound predicate below:

WHERE STATE = 'California'
AND CITY = 'Los Angeles'

Absent column correlation data, DB2 will estimate the number of qualifying rows by multiplying the cardinality of the STATE column by the cardinality of the CITY column; so, the estimate will be (1/50) * (1/50) * 10,000 rows = 4 rows. In fact, because STATE and CITY values are very highly correlated, the number of qualifying rows will be (1/50) * 10,000 = 200 (again, I'm assuming a uniform distribution of duplicate STATE and CITY values in the table's rows). Lack of correlation stats for the STATE and CITY columns results in a filtering estimate that is off by a factor of 50. When a filtering estimate is way off from reality, the access path chosen by the DB2 optimizer for a query could be sub-optimal, and that path may cause the query to run a lot longer than it needs to. KEYCARD gives DB2 more column correlation information to use in estimating the number of rows qualified by a query's predicates, and better information leads to better access path choices (and better query performance, which of course is the bottom line).

So, why not have RUNSTATS gather correlation stats for all combinations of all columns in a table? That could be done, using COLGROUP specifications for groups of non-indexed columns, but the CPU and run-time costs of doing this would be very high unless the table had very few columns and not many rows (plus, coding all the COLGROUP specifications for a table with a lot of columns would be extremely tedious). In most cases, it just isn't a practical option. Making KEYCARD automatic in DB2 10 is a good idea because it generates data that has significant query optimization value (additional column correlation information for indexes with keys comprised of more than two columns) and does so at a low cost in terms of RUNSTATS CPU and elapsed time (this because the column groups for which correlation data is obtained are "leading and contiguous" subsets of multi-column index keys, making the correlation data cheaply available by way of a table's indexes, in which key values are always strictly ordered).

I'll conclude with a word to the wise: don't wait for DB2 10 to make KEYCARD automatic in your environment. Go ahead and add this option to your RUNSTATS INDEX specifications. If, for example, your basic RUNSTATS control statement includes

TABLE(ALL) INDEX(ALL)

I'd recommend changing that to

TABLE(ALL) INDEX(ALL) KEYCARD

You'll enrich your catalog stats at a low incremental cost, and you'll have taken a step towards smoothing the path to DB2 10, as you'll have in your catalog the correlation stats that will be automatically generated in the DB2 10 environment.  

Thursday, December 2, 2010

REORG and DB2 for z/OS Partition-by-Growth Tablespaces

A DB2 DBA friend of mine recently told me of a problem he'd encountered in reorganizing a single partition of a partition-by-growth (PBG) tablespace: the online REORG job failed with an out-of-space condition pertaining to the partition's shadow data set. The topic of partition-level REORGs of PBG tablespaces is one that I find interesting; thus, this post, in which I'll explain why my friend hit that out-of-space situation and how he resolved it. I'll also provide some additional items of related information that I hope you'll find to be useful.

I blogged a couple of times about PBG tablespaces (introduced with DB2 9 for z/OS) while I was working as an independent DB2 consultant, in one entry comparing them with partition-by-range (PBR) tablespaces and in another briefly mentioning them among several new physical database design options provided by DB2 9. One advantage of a PBG tablespace versus a traditional segmented tablespace (PBG tablespaces are also segmented) is the ability to run utilities at a partition level (the exception to this rule being LOAD, which has to run at the tablespace level for a PBG tablespace). On the surface, the option of REORGing a single partition of a PBG tablespace looks pretty attractive, but there is a catch, and it does have to do with space, as my DBA friend discovered.

Here's the deal: by definition, a new partition for a PBG tablespace won't be allocated unless the tablespace's existing partitions are full. So, if you run an online REORG for, say, partition 4 of a ten-partition PBG tablespace, that partition is likely to be quite full (depending on delete activity, which might have created some "holes" in pages of the partition, and on the table's clustering key, which would influence placement of newly inserted rows -- more on that momentarily). If the tablespace was created with non-zero values for PCTFREE and/or FREEPAGE, the REORG utility will attempt to reestablish that free space, and if the partition's data rows and the reestablished free space will not fit in the partition's shadow data set, the REORG job will fail. That was the situation that my friend brought to my attention. What did he do? He resolved the problem by setting the value of a relatively new ZPARM parameter, REORG_IGNORE_FREESPACE (introduced last fall via the fix for APAR PK83397), to YES. When the value of REORG_IGNORE_FREESPACE is YES (the default is NO), REORG TABLESPACE will ignore PCTFREE and FREEPAGE values when reloading data rows into a PBG tablespace if either of the following is true:
  • The utility is reorganizing a subset of the tablespace's partitions
  • The table in the PBG tablespace has one or more LOB columns
If free space is not reestablished, the rows unloaded from a partition will almost certainly fit into the shadow data set during a REORG. [I say "almost" because for a table with varying-length rows (and rows in a compressed tablespace are varying-length even if there are no VARCHAR columns), I can imagine a very unusual scenario in which rows unloaded from a partition might not fit back in after being sorted in clustering key sequence, even with free space parameter values ignored. So, REORG_IGNORE_FREESPACE = YES should take chances of an out-of-space condition on REORG or a partition of a PBG tablespace way down, if not all the way to zero.]

Now, you might be thinking, "Why is there even the possibility of this out-of-space condition when one partition of a PBG tablespace is being reorganized? Can't REORG take rows that won't fit back into the target partition after reestablishing free space and put them in other partitions that have unused space?" The fact of the matter is that REORG doesn't work that way -- it doesn't move rows from one PBG partition to another unless the whole PBG tablespace is being REORGed or a range of the tablespace's partitions are being reorganized. To expand on that italicized point: if partitions 4 through 8 of my example ten-partition tablespace are being reorganized (i.e., if the utility control statement has PART 4:8), REORG can freely move rows across partitions 4, 5, 6, 7, and 8 (unless the DB2 version is 9 and the table in the tablespace has one or more LOB columns, in which case a row has to be reloaded into the partition from which it was unloaded -- a restriction removed with DB2 10): after sorting the rows unloaded from the five partitions in clustering sequence, REORG will fill one partition before moving to the next.

What this means: if you want to REORG a subset of a PBG tablespace's partitions, and you want free space to be reestablished by the utility, it may be better to REORG a range of several partitions instead of targeting a single partition. That way, unused space in one or more partitions in the range can enable restoration of free space in other, more-full partitions with less of a chance of out-of-space being an issue.

As a coda, I'll tell you that the DBA to whom I've referred repeatedly in this entry ended up asking whether it was even worth it to reorganize partitions of the table in question. That is a very good question to ask. It's easy to get into a REORG-if-it's-unclustered mindset, but in fact some tablespaces don't need to be REORGed. The one with which the DBA was dealing gets plenty of inserts but is rarely read, and when it is read, performance just has to be OK -- it doesn't have to be great. CPU resources in such a case can be saved by cutting way back on REORG frequency (if not eliminating REORGs altogether) and optimizing insert performance either by switching to a continuously-ascending clustering key or by altering the table with the APPEND YES option introduced with DB2 9 (this causes DB2 to ignore clustering when inserting or loading data into the table). Bottom line: don't REORG a tablespace (or partitions thereof) "just because." Make sure that the benefits of reorganizing an object justify the CPU (and disk space) used during REORG processing. While keeping an object well organized is usually a good idea, in some situations disorganization really isn't a big deal. Think it through, and act accordingly.