Friday, July 21, 2017

Db2 12 for z/OS SQL Enhancements: Result Set Pagination

In this second of a series of entries on my favorite SQL-related enhancements provided via Db2 12 for z/OS (and "Db2," versus "DB2," is officially the new form of the product name), I will describe useful functionality pertaining to result set pagination (my first post in this series covered piece-wise DELETE).

"Result set pagination" refers to retrieval of a query result set in "pages," with each page including a sub-set of the result set's rows. A typical scenario might involve presentation to a user of employee names (with, perhaps, accompanying information), starting with a particular name and going forward from there in ascending sequence based on last name followed by first name. Let's suppose that such a request has been initiated by a user, with the supplied "starting point" name being MARY ERIKSON.

Prior to Db2 12 for z/OS, the programmatic way of dealing with such a request was not entirely developer-friendly, in a couple of respects. First, there's the matter of generating the desired result set. That required a predicate with the following syntax:


A little clunky; and, it gets clunkier as more columns are examined (as would be the case if, for example, the request were to display names in ascending order by last name, then first name, then middle name, beginning with MARY TAYLOR ERIKSON).

The second not-totally-developer-friendly issue arises when the second page (and subsequent pages) of the result set has to be retrieved. Getting the first page is easy: you just fetch the first, say, 20 rows of the result set (if rows are to be displayed 20 to a page). How about getting the second page of 20 rows? Generally speaking, you could retrieve page n+1 of the result set by taking the last row in page n and plugging values from that row into a query predicate that would get you to the start of the next page's rows (and then fetch forward from there), utilizing the rather cumbersome syntax shown above. Referring to the example I've been using, if page 1 of the name-ordered result set ends with SAMUEL FIGGINS, page 2 could be retrieved via a query coded with the following predicate:


Again, this gets clunkier if more than two columns are to be considered for row ordering.

Alternatively, you could use ordinal positioning within the result set to generate succeeding pages, perhaps using a query with the OLAP specification ROW_NUMBER; or, you could just issue your initial result-set-building query, and fetch the first 40 rows and throw away the first 20 of those to display rows 21-40 in page 2 (an approach that gets progressively more wasteful, overhead-wise, as you progress through the result set's pages). Maybe you could use a scrollable cursor (though in some cases it might not be feasible to leave a cursor open for the purpose of retrieving succeeding pages of rows). In any case, you're looking at a level of coding complexity that is probably greater than you'd like.

Things get easier all around with Db2 12 (at function level V12R1M500 or later). For starters, initial result set generation is easier because of a newly supported predicate syntax. Want rows in last name, first name sequence, starting with ERIKSON, MARY? Just use this predicate:


And that simple syntax stays simple if more columns (e.g., middle name) are considered for row ordering.

Retrieval of succeeding result set pages also gets easier with Db2 12. You actually have a couple of alternative approaches, both simpler versus the pre-Db2 12 options. One option would be to take the last name of page n of the result set and use it to get to the starting point of page n+1, utilizing the simple syntax shown above. If, for example, FIGGINS, SAMUEL is in the last row of page 1 of our example result set, we can get to the starting point of page 2 by way of this predicate:


The second of the two new options also involves the use of SQL syntax previously unavailable with Db2 for z/OS: to get (keeping with the same example) the second page of 20 result set rows, just re-issue the query initially used to generate the result set, with this clause added (the new-with-Db2 12 part is underlined):


To get page 3, you'd issue the initial query with the clause OFFSET 40 ROWS FETCH FIRST 20 ROWS ONLY, and so on for subsequent pages, each time bumping the value "n" in OFFSET n ROWS up by 20, if you want 20 rows per page (and for the sake of consistency, you might issue the query used to get page 1 with the clause OFFSET 0 ROWS FETCH FIRST 20 ROWS ONLY). Note that both the OFFSET value and the FETCH FIRST value can be provided via variables, so the clause could look like this:


That way, the initial query can be used again and again as a user pages through the result set, with values provided as needed to get to the starting row of page n (and this approach also allows for the number of displayed rows to change, if desired, as pages beyond the first, or beyond the first few, are displayed).

Of the two new (with Db2 12) SQL pagination options, which should you use? The OFFSET n ROWS FETCH FIRST m ROWS approach delivers maximum coding simplicity. The other approach (plug values from the last row of page n into a query predicate to get to the starting point of page n+1, utilizing the simplified syntax shown previously) can provide a performance edge, as it takes you directly to the start of page n+1, whereas the OFFSET n ROWS option requires skipping over the first "n" rows of the initial result set -- an action that has some overhead cost (the performance difference could increase as the value of "n" in the OFFSET n ROWS clause gets larger). You can decide whether you want to go the maximally simple route, or the not-quite-as-simple-but-still-pretty-simple route that can be more efficient.

Whichever choice you make, it's a better one, from a developer's perspective, than the options available in a pre-Db2 12 environment.

More of my favorite Db2 12 SQL-related enhancements will be covered in subsequent posts to this blog, so check back in.

Friday, June 30, 2017

DB2 12 for z/OS SQL Enhancements: Piece-Wise DELETE

DB2 12 for z/OS, which became generally available in October of 2016, delivered a number of key enhancements pertaining to SQL statements. Over the course of a few blog entries, I'll describe some of my favorites among these enhancements. This first post in the series will cover piece-wise DELETE.

First, the backdrop: suppose you have a need to remove a large number of rows from a table, using SQL (I say, "using SQL," because a utility-based alternative, REORG with DISCARD, is not always feasible or desirable). How might you do that? Well, if the requirement is to remove all rows from table T1 that have a value greater than 100 in column C1, you could simply execute this statement:


Done, right? No so fast. How many rows are qualified by the C1 > 100 predicate? Suppose T1 has 500 million rows, and 50 million of those rows have a value greater than 100 in column C1? In that case, what happens when the DELETE statement shown above is executed? For one thing, the application process associated with the DELETE will potentially acquire a LOT of page or row locks (depending on whether LOCKSIZE PAGE or LOCKSIZE ROW is in effect for the table space in which T1 resides). That, in turn, could lead to the deleting program getting an error SQL code if the limit on the number of locks that can be acquired by one process (NUMLKUS in ZPARM) is reached. If that limit is not reached because lock escalation occurs (as it might, depending on the NUMLKTS value in ZPARM and the LOCKMAX setting in effect for T1's table space), other programs will not be able to access T1 until the DELETE statement completes and a COMMIT is issued by the deleting program. If the value of NUMLKUS in ZPARM is high enough so as to not be hit when the DELETE executes (or if NUMLKUS is set to 0, meaning there is no limit on the number of locks that can be held at one time by one process), and if lock escalation does not occur (because NUMLKTS is set to 0 or to a very high number, or if the table space's LOCKMAX value is 0 or a very high number), IRLM could run short on storage because of all the lock control blocks (and, in a DB2 data sharing environment, the lock list portion of the lock structure could fill up) -- not good for system stability. Oh, and one more thing: suppose that 50-million-row-impacting DELETE statement gets through 49 million rows and then encounters a problem that causes it to fail. What are you looking at then? One BIG rollback operation.

OK, so removing lots and lots of rows from a table with a single DELETE statement can be problematic. Is there a SQL alternative? Yes, but the alternative available in a pre-DB2 12 environment is not one about which an application developer would be enthused. I'm talking here about a cursor-based DELETE approach. That would involve writing a program that declares a cursor through which the to-be-deleted rows would be identified (using the preceding example, row identification would be done via a SELECT from T1 with the WHERE C1 > 100 predicate), with the DECLARE CURSOR statement including the WITH HOLD option to maintain cursor position through commits (the FOR UPDATE clause might also be required for the DECLARE CURSOR statement). The program would then OPEN the cursor, FETCH a row, and perform a DELETE WHERE CURRENT OF operation. Then FETCH another row, do another DELETE WHERE CURRENT OF, and so on, with commits performed periodically so as to release acquired locks, until all to-be-removed rows have been deleted from the target (a variant of this approach involves deleting blocks of rows using a cursor declared WITH ROWSET POSITIONING). Do-able, yes, but kind of clunky from a programming perspective. Wouldn't it be nice if there were a SQL mechanism through which large-scale DELETEs could be accomplished in a way that would be programmer-friendly while at the same time avoiding the problems associated with massive accumulation of locks (or with lock escalation)? Well, thanks to DB2 12 for z/OS (at function level V12R1M500 or above), that mechanism is here. It's called piece-wise DELETE.

Using piece-wise DELETE, how easy is it to perform a large-scale row-removal operation that does not gum up the works, concurrency-wise? This easy (and here I'll assume that we want to remove rows from the target table in chunks of 500):

  2. Issue a COMMIT
  3. Repeat steps 1 and 2 until all to-be-removed rows have been deleted
The part of the DELETE statement highlighted in red above is syntax that's newly supported with DB2 12 (again, at function level V12R1M500 or above): FETCH FIRST n ROWS ONLY can be coded in a DELETE statement, and doing that makes the DELETE of the piece-wise variety. Every execution of the statement (as coded above) will remove 500 rows from the target table.

What if you sometimes want to use your piece-wise DELETE program to remove rows in chunks of 50, or 1000, instead of 500? No prob. You can use a variable to specify the to-be-deleted quantity in the DELETE statement's FETCH FIRST clause. And, if you're using a variable to provide the FETCH FIRST value, you can change the row-removal "chunk" size as a large-scale DELETE operation is executing.

There you have it -- an easy-to-program means of breaking up a DELETE operation into bite-sized pieces (or bite-sized units of work, in more technical parlance). In blog posts to come I'll cover others of my favorite DB2 12 SQL enhancements. I hope you'll check those out.

Wednesday, May 24, 2017

DB2 for z/OS and Data-as-a-Service and Database-as-a-Service

______-as-a-service is all the rage in IT these days, and understandably so -- the "service-ization" of information technology capabilities and interfaces will be transformational; indeed, transformations have already occurred or are underway at many organizations and within many application development and deployment environments. In this blog entry I want to highlight a couple of ______-as-a-service concepts, data-as-a-service (aka DaaS) and database-as-a-service (aka DBaaS), and examine their relevance to DB2 for z/OS. First, a little level-setting terminology clarification is in order.

Data-as-a-service and database-as-a-service -- NOT the same thing

Some folks use the terms "data-as-a-service" and "database-as-a-service" interchangeably, but they do not have the same meaning. The distinction is indicated by the different first words of the terms. When your're talking database-as-a-service, you're talking about the functionality of a database management system, provided as a service. So, what does that mean? Well, it can mean that the database management system in question is deployed in an off-premise cloud environment, but database-as-a-service can also apply to a DBMS that is deployed in an on-premise fashion, and that brings me to what I feel is a very important point: "cloud" doesn't necessarily mean "off-premise," and "off-premise" doesn't necessarily mean "cloud." Take the latter part of this point I just made. For a LONG time, numerous organizations have contracted with various companies (IBM among them) to manage their IT operations. IT outsourcing company X might run organization Y's IT operations out of a data center owned by X. That's off-premise from Y's perspective, but there may be nothing cloudy about the way in which Y's IT operations are managed by X. Similarly, an organization could manage some or all of its IT resources in a cloudy way, even with all the IT resources managed in that way being on-premise. Cloud is about consumption of ______ as a service. To those who consume the service(s) provided, the location of the mechanism of service provision is (or should be) irrelevant.

OK, if database-as-a-service can be implemented in an on-premise way (and again, it can), what kind of services are we talking about, and who consumes those services? Quite often, the consumers are application developers, and the services provided are things like a new schema, or a new database instance, or a data backup or restore operation. More on this a little later on.

While database-as-a-service is concerned with the requisitioning and provisioning of database management system functionality, data-as-a-service is about the programmatic interface to a data server. "Database" is not part of the term because a user, who could be a developer writing a client application, has no need to know (and likely has no desire to know) whether or not a database management system is on the server side of a data request. Many application developers -- even those with strong SQL skills -- see the mechanism by which a data request is serviced (which could be a relational database management system such as DB2 for z/OS) as being part of back-end "plumbing," and plumbing is not their concern. They want a straightforward and consistent interface through which CRUD data services (create, read, update, delete) can be invoked -- an interface that abstracts virtually all of the particulars of the service-providing system. REST (short for REpresentational State Transfer) is one such interface, and in recent years it has become a very popular means of enabling data-as-a-service application architecture.

Data-as-a-service in a DB2 for z/OS context

Responding to the growing preference on the part of application developers for as-a-service interaction with data-providing systems, IBM built a native REST interface into DB2 12 for z/OS, and retrofitted that interface to DB2 11 (via the fix for APAR PI66828). DB2's native REST interface is an extension of the DB2 distributed data facility (DDF), so it leverages existing DDF capabilities such as thread pooling, profile monitoring, workload classification, and accounting and statistics tracing. Another benefit of the DDF tie-in: a SQL statement invoked by way of DB2's REST interface will execute under a preemptible SRB in the DDF address space, making the statement zIIP-eligible (up to 60%).

By way of DB2's REST interface, a single static SQL statement can be exposed in the form of a REST API. Before you start thinking of that "single statement" thing as overly limiting, consider that the single statement can be a call to a DB2 stored procedure (if you decide to make a stored procedure a RESTful service, I'd recommend going with a native SQL procedure, because a native SQL procedure is up to 60% zIIP-eligible when it executes under a DDF preemptible SRB).

In addition to enabling invocation of SQL statements via REST calls, DB2's REST-related capabilities support creation, discovery, management, and securing of SQL-based RESTful services. That functionality becomes more extensive (and more user-friendly) when z/OS Connect front-ends DB2's REST interface (i.e., when DB2 is a REST provider to z/OS Connect); and, z/OS Connect makes all kinds of z/OS-based programmatic assets available to REST clients: not only SQL statements (and stored procedures), but CICS transactions, IMS transactions, WebSphere Application Server for z/OS transactions, and batch jobs, as well.

Database-as-a-service in a DB2 for z/OS context

Quite a lot of work has been going on behind the scenes to provide database-as-a-service capabilities for DB2 for z/OS, and the first fruits of those efforts recently showed up in the form of a pair of DB2 APARs -- more on that in a moment.

Recall my mentioning, early on in this blog entry, that a raison d'etre for an on-premise database-as-a-service implementation could be easy and speedy requisitioning and provisioning of DBMS functions to support the work of application developers. The team within IBM DB2 for z/OS development that is leading the database-as-a-service charge has been focused on precisely that. And so it is that we have new DB2 for z/OS database-as-a-service capabilities that have recently been introduced via the fixes for APARs PI73168 (for DB2 11) and PI73492 (for DB2 12). These APARs leverage the functionality of zOSMF (the z/OS Management Facility), a component of the z/OS operating system that can drive the execution of "workflows" (roughly analogous to what are referred to as "scripts" in the distributed systems world) in response to REST calls. The APARs make available zOSMF workflows that can provide three database functions in an as-a-service manner: creation of a new database, creation of a new schema, and creation of a copy of an existing schema. The latter function, referred to as "schema like" (as in, "create a new schema like that one"), involves using, in addition to z/OSMF, the capabilities of the IBM DB2 Change Management Solution Pack for z/OS, which itself combines the functionality of the IBM DB2 Administration Tool and the IBM Object Comparison Tool.

More DB2 for z/OS database-as-a-service capabilities are being worked on, so watch this space.

All of this is good news for people who support DB2 for z/OS. Making data managed by DB2, and making database functionality provided by DB2, more easily consumable will help development teams to be more agile and more productive in building applications that interact with DB2. Additionally, more fully automating the requisitioning and provisioning of DB2 database management system resources will free DBAs from lots of "blocking and tackling" work around things like setting up new databases and schemas in development environments, and that should enable them to spend more time engaged in higher-value activities like working with developers on the design and architecture of new DB2-accessing applications.

Sunday, April 30, 2017

DB2 for z/OS: Something You Might Not Know About Large Page Frames in a Post-DB2 10 Environment

I well remember the first time I saw it. August, 2016. I was looking at output from the DB2 for z/OS command -DISPLAY BUFFERPOOL(ACTIVE) DETAIL, issued on a DB2 11 system, that a client had sent to me for review. I appreciated the fact that I could see in this command output information about DB2's use of large real storage page frames for page-fixed buffer pools (in a DB2 10 environment, one had to issue -DISPLAY BUFFERPOOL -- for an individual pool, versus for all active pools -- with an odd syntax to get this information). I saw pools for which PGFIX(YES) had been specified, and observed that, as expected, DB2 showed that the preferred real storage frame size for these pools was 1 MB. Then I noticed something I couldn't explain: there was a buffer pool, with a PGFIX(YES) specification, and DB2 was telling me, 


Huh? Why would DB2 prefer 4 KB-sized page frames for that pool? PGFIX(YES) buffer pools are supposed to be backed by 1 MB page frames, right? The indicated preference for 4 KB page frames was not the result of the pool being smaller than 1 MB: the pool's VPSIZE value was several thousand, and all it takes is 256 buffers of 4 KB apiece to fill a 1 MB frame. I continued through the -DISPLAY BUFFERPOOL output, and got even more confused. There were some PGFIX(YES) pools for which 1 MB was the indicated page frame size preference, and other PGFIX(YES) pools for which 4 KB was seen to be the preferred real storage page frame size. There did not appear to me to be any rhyme or reason for this variance. I moved on, and ultimately forgot about the strange-looking DISPLAY BUFFERPOOL output.

Fast forward about seven months, and I'm talking to a group of people about DB2 for z/OS buffer pools and large real storage page frames. I told my audience that DB2 will automatically seek to allocate a PGFIX(YES) buffer pool using 1 MB page frames. "That's not true," said a person in the room. I was pretty sure that this guy was wrong on that point, but in the interest of time I told him, "Let's talk about this offline." Over the next couple of days, this individual and I exchanged e-mails on the topic, and he made a very good argument in support of his contention and I tried some things on an IBM-internal DB2 for z/OS subsystem and, by golly, the man was right.

Here's the deal: in a DB2 10 for z/OS environment, 1 MB page frames will be preferred for a buffer pools for which PGFIX(YES) has been specified. In a DB2 11 (or later) system, that may be the case. Why "may," instead of "will?" Because DB2 11 introduced the FRAMESIZE option for the -ALTER BUFFERPOOL command. That option lets you tell DB2 of your preference for page frame size for a buffer pool, and for a PGFIX(YES) pool your preference becomes DB2's preference (assuming that there are enough buffers, per the pool's VPSIZE value, to fill at least one of the page frames of the size indicated via the FRAMESIZE specification). As DB2 10 did not have a FRAMESIZE option for ALTER BUFFERPOOL, that specification was, in essence, implicitly conjoined with the PGFIX specification -- when you requested, in a DB2 10 system, that a pool's buffers be fixed in the z/OS LPAR's real storage, you were also implicitly requesting allocation of the pool's buffers in 1 MB real storage page frames (subject to the availability of such page frames in the LPAR).

DB2 11 for z/OS introduced a choice for real storage page frame size preference for a page-fixed buffer pool: 1 MB or 2 GB. The FRAMESIZE option was added to the -ALTER BUFFERPOOL command to enable explicit specification of the preferred page frame size for a pool. It was made pretty clear that use of 2 GB page frames for a buffer pool required a FRAMESIZE(2G) specification for the pool. Lots of people (myself included) assumed that specifying FRAMESIZE(1M) was not required if one wanted a page-fixed buffer pool to be allocated using 1 MB page frames. We in this camp believed that FRAMESIZE(1M) was automatically specified "under the covers" in a DB2 11 system when an -ALTER BUFFERPOOL command with PGFIX(YES) was issued. As it turns out, that was a FALSE ASSUMPTION on our part. The facts are these: first, in a DB2 11 (or later) system, 1 MB will be the preferred real storage page frame size, even without a FRAMESIZE(1M) request, for a pool that was page-fixed and in-use in the DB2 10 environment, prior to the migration of the system to DB2 11. Second, for a pool newly allocated in a DB2 11 (or later) environment, an explicit FRAMESIZE(1M) specification is required if one wants 1 MB to be the preferred real storage page frame size for a pool.

To put this another way: as part of the migration of a DB2 10 system to DB2 11, existing page-fixed buffer pools were "migrated" to the DB2 11 environment with an automatic FRAMESIZE(1M) specification added to the pools' definition. That makes sense, because it preserves the behavior of the pools in the DB2 11 environment that was in effect in the DB2 10 system. For new pools allocated in a DB2 11 (or later) system, because requesting a page frame size preference is a separate action from requesting page-fixed buffers, you have to explicitly specify FRAMESIZE(1M) or FRAMESIZE(2G) in order make the preferred frame size 1 MB or 2 GB, respectively.

Here are your take-aways:

  1. If your DB2 for z/OS environment is at Version 11 (or later), issue the command DISPLAY BUFFERPOOL(ACTIVE) DETAIL, and examine the output for each pool. Do you see any pools for which the PGFIX attribute is YES, and the indicated PREFERRED FRAME SIZE is 4K? If yes, it is likely that those pools were in use before the system was migrated to DB2 11. If you want 1 MB page frames to be used for the pools, alter them with a FRAMESIZE(1M) specification.
  2. Even if all of the PREFERRED FRAME SIZE attributes for your buffer pools are as you want them to be, give some thought to altering each pool (or at least each pool for which PGFIX(YES) is an attribute) with the appropriate FRAMESIZE specification. I think that there is value in having an explicit FRAMESIZE specification for each of your buffer pools (or, again, at least each of the page-fixed pools), even if that explicit frame size preference is the same as the one that is in effect by default for a pool. This serves to get you into the "make it clear by making it explicit" mind set, and that can be a sound aspect of your buffer pool management strategy.
As always, thanks for visiting the blog.

Sunday, March 26, 2017

DB2 for z/OS: Running REORG to Reclaim Disk Space

Think of why you run the DB2 for z/OS REORG utility, and a number of reasons are likely to come quickly to mind: to restore row order per a table's clustering key; to reestablish free space (for inserts and/or for updates); to remove the AREO* status set for a table space following (for example) an ALTER TABLE ADD COLUMN operation; or to materialize a pending DDL change such as an enlargement of a table space's DSSIZE. How about disk space reclamation? If that REORG motivation has not previously occurred to you, perhaps it should.

Recently, a DBA at a large DB2 for z/OS site communicated to me the success that his team has had in reclaiming substantial amounts of disk space through online reorganization of certain table spaces. He also asked for a recommendation with regard to identifying table spaces for which a REORG could potentially deliver significantly reduced disk space consumption. In this blog entry, I'll describe the disk space reclamation scenario reported by the referenced DBA, I'll explain why there was space to be reclaimed in some of the table spaces administered by the DBA, and I'll provide the "reclamation indicator" metric that I suggested to the DBA as a means of identifying table spaces that could be reorganized in order to free up disk space.

First, the scenario. At the DBA's site, there are some tables, in segmented table spaces ("traditional" segmented table spaces, as opposed to universal table spaces, which also happen to be segmented), that have these key characteristics: they are clustered by a continuously-ascending key (so that "new" rows go to the "end" of the table), and the number of inserts into the table is roughly equaled by the number of rows that are deleted from the table over a period of time.

The DB2 DBA knew that for table spaces with the above-described characteristics, REORGs were not needed to maintain "clusteredness," because of the continuously-ascending clustering key that sent new rows to the end of the table (at least, clustering would remain in good shape until the table space reached its size limit -- more on this in a moment). For the same reason, free space for inserts in "interior" pages of the table space was not a concern. Still, with DB2 real-time statistics showing a very large number of inserts since the last REORG of a couple of these table spaces, the DBA determined that reorganizations might be in order. Online REORGs of the table spaces were executed, and the result was a freeing up of 64 GB of disk space: one table space went from 21 to 4 data sets of 2 GB apiece, and the other went from 17 data sets to 2 (a DB2 segmented table space is comprised of up to 32 data sets of 2 GB apiece, and that is why its size limit is 64 GB).

Why was there so much unused space in these table spaces? Because the continuously-ascending clustering key kept pushing the "end" of the table spaces "outward." Why would that happen? Why would DB2 grow these table spaces as a result of inserts, given the like number of row-delete operations that were targeting the associated tables? Shouldn't DB2 have been using the space freed up by deletes to accommodate new inserts, without growing the table space's size? Actually, DB2 was working as designed. It's true that, given a continuously-ascending clustering key and some deletes of older rows from the "front" of a table space, DB2 can "wrap" to the front and start inserting new rows in space cleared by deletes, but that will only happen if DB2 cannot extend the table space (i.e., if DB2 cannot make the table space larger). If DB2 can extend a segmented table space, it will in order to preserve a table's row-clustering order; so, in advance of hitting the 64 GB size limit for a segmented table space, DB2 would keep making the table space larger so that it could keep adding rows to the end of a table (assuming a continuously-ascending clustering key), and deletes of older rows would result in ever-larger amounts of available-but-unused space in the table space. That's why the disk footprint of the two table spaces became so much smaller following reorganization.

[It is important to keep in mind that, given a continuously-ascending clustering key and at least some row-delete operations, DB2 will insert new rows in the "front" of a segmented table space, using space freed up by DELETEs, if the table space cannot be made any larger (either because of reaching the 64 GB limit or as a result of running into a maximum-extents or a maximum-volumes situation). In that case, "wrapping to the front" for new inserts is better than failing the inserts.]

Having achieved this disk space reclamation success through REORGs, the aforementioned DBA wanted to add "potential for significant disk space reclamation" to the criteria used at his site for identifying table spaces that should be reorganized (a good proactice -- REORG table spaces when you have a good reason for doing so, not just "because it's been X amount of time since the last time this table space was REORGed"). How could he and his colleagues spot table spaces with large amounts of unused space? My recommendation: use for this purpose the ratio of disk space occupied by the table space to space in the table space occupied by rows in the table space. For the numerator, I'd use the SPACE value in the row for the table space in the SYSTABLESPACE catalog table. That value is updated when the STOSPACE utility is executed, so you would want to run STOSPACE on a regular basis (that should not be a big deal -- STOSPACE should be a very inexpensive utility to execute, CPU-wise). For the denominator, I would use the product of TOTALROWS from SYSTABLESPACESTATS (set by REORG and updated when INSERTs and DELETEs are executed) and AVGROWLEN in SYSTABLESPACE (updated by RUNSTATS, or by in-line statistics collected during REORG or LOAD). You can decide when that ratio would prompt you to run REORG to reclaim space. Would you do that when disk-space-to-row-space hits 2 (i.e., when the size of the table space is 2X the space occupied by rows)? When it hits 3? When it hits 4? One of those values might be reasonable for your environment.

One more thing: I have focused on traditional segmented table spaces in this blog entry because that is the table space type to which space reclamation via REORG is most relevant. For a range-partitioned table space, a given partition's size limit is determined by the DSSIZE specification, and the same is true for a partition-by-growth table space. Yes, you could see a partition-by-growth table space come to contain a high percentage of unused space given the combination of a continuously-ascending clustering key and a good deal of DELETE activity, but you could put a limit on that growth by way of a not-larger-than-needed MAXPARTITIONS value. With that said, even with range-partitioned and partition-by-growth table spaces you could see situations in which the ratio of table space size to space occupied by rows (the ratio described in the preceding paragraph) could get to be high enough to make a REORG attractive from a disk space reclamation perspective. And here there's some good news: starting with DB2 11 for z/OS, you can tell DB2 to drop partitions of a partition-by-growth table space made empty by a REORG or the entire table space (that functionality is enabled via the REORG_DROP_PBG_PARTS parameter in ZPARM).

So, add disk space reclamation to your reasons for running REORG (if you have not already done so), and consider using the ratio I've provided to look for candidate table spaces.

Sunday, February 26, 2017

DB2 for z/OS: the PGSTEAL and PGFIX Options of -ALTER BUFFERPOOL

Recently, a DB2 for z/OS professional I’ve known for some years sent to me a question about the relationship between the PGSTEAL and PGFIX options of the DB2 command -ALTER BUFFERPOOL. It took a few iterations of question and answer to get things straightened out, and I thought, “Hmm. If this person, who has lots of DB2 for z/OS experience and knowledge, needed a little help in getting PGSTEAL and PGFIX straightened out in his mind, it's likely that other DB2 people are in the same boat.” And so I’m writing this blog entry, in the hope that it will be helpful to people who have some uncertainty about the dependencies – if any – between PGSTEAL and PGFIX.

OK, the “if any” in the preceding sentence might suggest to you that maybe there isn’t anything in the way of interdependency with regard to PGSTEAL and PGFIX. In fact, there really isn't. Yes, there’s a recommended combination of PGSTEAL and PGFIX settings, and I’ll get to that, but the suggested combination is about the way in which PGSTEAL and PGFIX can both support a particular performance objective, as opposed to having anything to do with the way these two buffer pool specifications affect each other, because they don't affect each other.

“What?” you might ask. “How is it that two buffer pool configuration options that both have 'page' in their long names (‘PG’ is short for ‘page’) don’t really have anything to do with each other?” This is so because PGSTEAL and PGFIX have very different functions. This functional difference might have been more readily apparent to people if PGSTEAL had instead been labeled something like BSTEAL – short for “buffer steal,” because PGSTEAL is about management of buffers, which is a DB2 responsibility, and PGFIX is about management of real storage – a responsibility of the z/OS operating system.

Let me make this distinction even more clear: PGSTEAL is related to virtual storage management – DB2’s management of virtual storage space that belongs to DB2 – and PGFIX is related to real storage management – something in z/OS’s bailiwick.

Consider the possible values of PGSTEAL: LRU, FIFO, and NONE. The first two are specified when you anticipate that there will be some buffer-steal activity for the pool in question (i.e., there are more pages in objects assigned to the pool than there are buffers in the pool). You’d go with LRU (least recently used) if you anticipate that there will be quite a bit of buffer-steal activity for the pool (“OK, DB2, you’re probably going to have to frequently steal buffers for this pool. When you do that – when you have to replace a table or index page currently in a buffer with another page – steal the buffer holding the page that’s gone the longest time without being referenced.”). You’d choose FIFO (first in, first out) when you expect some – but very little – buffer-stealing for the pool (“Hey, DB2. You’re going to have to do a very small amount of buffer stealing for this pool. That being the case, don’t waste CPU cycles keeping track of which buffers hold the pages that have gone the longest time without being referenced. Just steal the buffer holding the page brought into memory the longest time ago.”).

How about PGSTEAL(NONE)? When would you go with that option? You’d go this route if you anticipate that there will be NO buffer-steal activity for a pool – in other words, the pool has more buffers than there are pages belonging to objects assigned to the pool (“Hey, DB2. I’m planning on using this pool to cache one or more table spaces and/or indexes in memory in their entirety. I’m doing that because I want maximum performance for these babies. Help me out.”). And, given the message you’ve sent to DB2, DB2 will help you out: it will asynchronously read into the pool every page belonging to an object assigned to the PGSTEAL(NONE) pool when the object is first referenced after a DB2 start-up, and in optimizing SQL statements targeting objects assigned to the pool, it will perform access path selection with the assumption that read I/O costs will be zero (not counting possible reads associated with use of work file table spaces).

Next, PGFIX settings. There are two: YES and NO. When NO (the default) is in effect, a real storage page frame occupied by a DB2 buffer can be stolen by z/OS (i.e., the contents of the page frame can be moved to auxiliary storage, also known as the page data sets) in order to make room for some other page that has to be brought into real storage (in performing such page steal actions, z/OS utilizes a least-recently-used algorithm). To prevent a z/OS real storage page frame steal action from interfering with a buffer read I/O (DB2 copies a table space or index page from disk – or maybe from a group buffer pool, in a data sharing system – into a buffer) or write I/O operation (DB2 copies a page in a buffer to disk or to a group buffer pool), a buffer will be fixed in memory (i.e., made non-page-able) prior to the I/O action, and released (made page-able again) after the I/O action. When PGFIX(YES) is in effect for a pool, buffers belonging to the pool are fixed in memory (made non-page-able) as soon as the pool is allocated, and they remain fixed in memory as long as the pool is allocated. Because the buffers are always in a fixed-in-memory state, the page-fix/page-release actions required for all buffer read and write I/O operations when PGFIX(NO) is used are not necessary. That makes buffer read and write I/Os more CPU-efficient when PGFIX is set to YES, and that makes PGFIX(YES) particularly beneficial for pools with high rates of read and write I/Os.

PGFIX(YES) is also a prerequisite if you want a DB2 buffer pool to be backed by large real storage page frames (1 MB or, starting with DB2 11 for z/OS, 2 GB frames, versus the traditional 4 KB frames). When a pool’s buffers are located in large real storage page frames, further CPU savings (beyond less-expensive I/O operations) are realized, because large real storage page frames make virtual-storage-to-real-storage address translation more efficient.

If we take a look at some combinations of PGSTEAL and PGFIX settings, you’ll see what I meant about there being essentially nothing in the way of interdependency. Take PGSTEAL(LRU) and PGFIX(YES). Will the fact that the pool’s buffers are fixed in memory have any impact on the level of buffer stealing done by DB2 for this pool? NO! REGARDLESS of whether a pool’s buffers are fixed in memory or not, DB2 will ALWAYS steal a buffer in the pool WHENEVER IT HAS TO, and a buffer HAS TO BE STOLEN whenever all the pool’s buffers are occupied (by table space and/or index pages) and a page has to be read into the pool from disk (or from a group buffer pool in a coupling facility). The PGFIX(YES) setting for this pool simply means that z/OS will not move any of the pool’s buffers out of real storage to auxiliary storage – it in no way restricts DB2’s ability to steal buffers so that pages not yet in the pool can be brought in from disk (or from a coupling facility).

Change the combination of PGSTEAL and PGFIX settings to NONE and NO, respectively, and again you’ll see that there is nothing in the way of an interdependency. PGSTEAL(NONE) means that DB2 does not EXPECT to have to do any buffer stealing for the pool, because there is an assumption that the number of pages belonging to objects assigned to the pool does not exceed the quantity of buffers in the pool (i.e., does not exceed the pool’s VPSIZE). Of course, if the number of pages in objects assigned to the pool does exceed the pool’s VPSIZE (perhaps VPSIZE was too small to begin with, or it was OK initially but objects assigned to the pool have gotten larger), and a buffer has to be stolen because all buffers are occupied and a page has to be brought into the pool, a buffer WILL BE STOLEN because (as noted previously) a buffer will ALWAYS BE STOLEN when a buffer HAS TO BE STOLEN (when PGSTEAL is set to NONE for a pool and some buffer stealing has to be performed, it will be done on a FIFO basis – first in, first out). What about the PGFIX(NO) setting for this pool? Will the pool’s PGSTEAL(NONE) setting in any way restrict the ability of z/OS to move buffers belonging to the pool out of real storage to auxiliary storage, as necessary, to free up real storage page frames so that new pages can be brought into real storage? NO! With PGFIX(NO) in effect, a pool’s buffers are absolutely fair game for being paged out to auxiliary storage, REGARDLESS of the pool’s PGSTEAL setting (that said, buffers in PGFIX(NO) pools are not often paged out of real storage, because z/OS, as noted, steals real storage page frames on a least-recently-used basis, and DB2 tends to “touch” its buffers with a frequency that makes them unlikely to be among the least-recently-used pages in a z/OS LPAR’s real storage).

DB2, then, manages its buffers as it needs to (with some guidance provided by you via a pool’s PGSTEAL setting), and z/OS manages its real storage resource as it needs to (with the understanding that buffers in a PGFIX(YES) pool are off-limits with regard to page-out actions), and the DBMS and the OS attend to these respective tasks pretty much independently. PGSTEAL settings do not impact z/OS page-frame-stealing, and PGFIX settings do not impact DB2 buffet-stealing.

Now, early on in this post I mentioned that while there essentially aren’t any interdependencies between the various settings of PGSTEAL and PGFIX for a buffer pool, there is a recommendation I have concerning a particular combination of PGSTEAL and PGFIX values. Here it is: if you are going to specify PGSTEAL(NONE) for a buffer pool, specify PGFIX(YES) for that same pool, unless the demand paging rate of the z/OS LPAR is higher than it should be (if the demand paging rate – available by way of a z/OS monitor – is in the low single digits or less per second, it’s OK). Why I make this recommendation: presumably, you assign objects to a PGSTEAL(NONE) buffer pool because you in fact want them to be cached in memory in their entirety. You would do this, I imagine, for objects for which maximum performance is really important. If that’s the case, why not really max out performance as it pertains to accessing these objects? Make the pool PGFIX(YES) as well as PGSTEAL(NONE), so that you can get the CPU efficiency benefit of large real storage page frames (of course, to get that benefit you need to have large page frames available in the LPAR to back the PGFIX(YES) pool – information about that can be found in an entry I posted to this blog a fewmonths ago).

And there you have it. Just remember that PGSTEAL is related to DB2’s management of its buffers, while PGFIX is related to z/OS’s management of the LPAR’s real storage resource. Two different buffer pool configuration settings, for two different aspects of DB2 performance.

Tuesday, January 31, 2017

Are You Using System Profile Monitoring to Manage Your DB2 for z/OS DDF Workload? Perhaps You Should

Here's a scenario that might sound familiar to you: you have a particular transaction, which I'll call TRNX, that is the source of quite a lot of deadlock activity in a DB2 for z/OS system. It seems that whenever more than one instance of TRNX is executing at the same time, a deadlock situation is highly likely. You went with row-level locking for the table spaces accessed by TRNX, but the trouble persisted. It is conceivable that rewriting the program code associated with TRNX might eliminate the problem, but the task would be challenging, the development team has limited bandwidth to accomplish the recommended modifications, and it could take months -- or longer -- for the fix to get into production. What can you do?

Well, as I pointed out in an entry posted to this blog a few years ago, sometimes the right approach in a case such as this one is to single-thread TRNX. Though it may at first seem counter-intuitive, there are circumstances for which transactional throughput can be increased through a decrease in the degree of transactional multi-threading, and that approach can be particularly effective when the rate of transaction arrival is not particularly high (i.e., not hundreds or thousands per second), transaction elapsed time is short (ideally, well under a second), and probability of a DB2 deadlock is high if more than one instance of the transaction is executing at the same time.

Lots of people know how to single thread a CICS-DB2 or IMS-DB2 transaction, but what about a DDF transaction (i.e., a transaction associated with a DRDA requester, which would be an application that accesses DB2 for z/OS by way of TCP/IP network connections)? Is there a means by which a DDF transaction can be single-threaded?

The answer to that question is, "Yes," and the means is called system profile monitoring, and DDF transaction single-threading is just one of many useful applications of this DB2 for z/OS capability. I'll provide a general overview of DB2 system profile monitoring, and then I will cover transaction single-threading and a couple of other use cases.

System profile monitoring is put into effect by way of two DB2 tables, SYSIBM.DSN_PROFILE_TABLE, and SYSIBM.DSN_PROFILE_ATTRIBUTES. Those tables were introduced with DB2 9 for z/OS, and DB2 10 enabled their use in managing a DDF workload in a more granular fashion than was previously possible. Prior to DB2 10, the number of connections from DRDA requesters allowed for a DB2 subsystem, and the maximum number of those connections that could be concurrently in-use, and the maximum time that an in-use (i.e., non-pooled) DBAT (database access thread -- in other words, a DDF thread) could sit idle without being timed out, could be controlled only at the DB2 subsystem level via the ZPARM parameters CONDBAT, MAXDBAT, and IDTHTOIN, respectively. What if you want to exert control over a part of a DDF workload in a very specific way? With system profile monitoring, that is not at all hard to do.

A row inserted into SYSIBM.DSN_PROFILE_TABLE indicates the scope of a particular DDF workload managemnt action, and the corresponding row (or rows) in SYSIBM.DSN_PROFILE_ATTRIBUTES indicates what you want to manage for this component of your DDF workload (number of connections, number of active connections, idle thread timeout, or two of the three or all three) and how you want that management function to be effected (e.g., do you want DB2 to take action when a specified limit is exceeded, or just issue a warning message). The columns of the two tables, and their function and allowable values, are well described in the DB2 for z/OS documentation, and I won't repeat all that information here (the DB2 11 information is available online at, and you can easily go from there to the DB2 10 or DB2 12 information, if you'd like). What I will do is take you through a few use cases, starting with the single-threading scenario previously referenced.

To single-thread a DDF transaction, you would first need to identify that transaction by way of a row inserted into the SYSIBM.DSN_PROFILE_TABLE. You have multiple options here. You might identify the transaction by workstation name (a string that is easily set-able on the client side of a DDF-using application, as described in a blog entry I wrote back in 2014); or, you might identify the transaction via package name, if, for example, it involves execution of a particular stored procedure; or, you might use collection name [Collection name can be specified as a client-side data source property, and it is increasingly used to manage applications that use only IBM Data Server Driver (or DB2 Connect) packages -- these packages, which belong by default in the NULLID collection, can be copied into other collections, and in that way a DDF-using application can be singled out by way of the name of the Data Server Driver (or DB2 Connect) package collection to which it is pointed.] And, there are multiple additional identifier choices available to you -- check the DB2 documentation to which I provided the link above.

In your SYSIBM.DSN_PROFILE_TABLE row used to identify the transaction you want to single-thread, you provide a profile ID. That ID serves as the link to an associated row (or rows) in SYSIBM.DSN_PROFILE_ATTRIBUTES. In a row in that latter table, you would provide the ID of the profile you'll use to single-thread the transaction, 'MONITOR THREADS' in the KEYWORD column, 1 in the ATTRIBUTE2 column (to show that you will allow one active DBAT for the identified transaction), and 'EXCEPTION' in the ATTRIBUTE1 column to indicate that DB2 is to enforce the limit you've specified, as opposed to merely issuing a warning message (you could also specify 'EXCEPTION_DIAGLEVEL2' if you'd like the console message issued by DB2 in the event of an exceeded threshold to be more detailed versus the message issued with EXCEPTION, or its equivalent, EXCEPTION_DIAGLEVEL1, in effect). Then you'd activate the profile with the DB2 command -START PROFILE, and bingo -- you have put single-threading in effect for the DDF transaction in question.

Something to note here: Suppose you have set up single-threading in this way for transaction TRNX, and an instance of TRNX is executing, using the one thread you've made available for the transaction. Suppose another request to execute TRNX arrives. What then? In that case, the second-on-the-scene request for TRNX will be queued until the first-arriving TRNX completes (if TRNX typically executes in, say, less than a second, the wait shouldn't be long). What if a third request for TRNX comes in, while the second request is still queued because the first TRNX has not yet completed? In that case, the third TRNX request will fail with a -30041 SQLCODE. This is so because DB2 will queue requests only up to the value of the threshold specified. If you specify 1 active thread for a transaction, DB2 will queue up to 1 request for that transaction. If you specify a maximum of 4 active threads for a transaction, DB2 will queue up to 4 requests for the transaction if the 4 allowable active threads are busy. With this in mind, you'd want to have the TRNX program code handle the -30041 SQLCODE and retry the request in the event of that SQLCODE being received. Would you like to be able to request a "queue depth" that is greater than your specified threshold value? So would some other folks. That enhancement request has been communicated to the DB2 for z/OS development team.

Something else to note here: What if you are running DB2 in data sharing mode. Does a limit specified via SYSIBM.DSN_PROFILE_TABLE and SYSIBM.DSN_PROFILE_ATTRIBUTES apply to the whole data sharing group? No -- it applies to each member of the group. How, then, could you truly single-thread a DDF transaction in a data sharing environment? Not too hard. You'd set up the profile and associated threshold as described above, and you'd start the profile on just one member of the group (-START PROFILE is a member-scope command). On the client side, you'd have the application associated with TRNX connect to a location alias, versus connecting to the group's location, and that alias would map to the one member for which the single-threading profile has been started (I wrote about location aliases in an entry posted to this blog a few years ago -- they are easy to set up and change). If the one member is down, have the profile ready to go for another member of the group (you could have leave the GROUP_MEMBER column blank in the DSN_PROFILE_TABLE row to show that the profile applies to all members of the group, or you could have two rows, one for the "primary" member for the transaction in question, and one for an alternate member, in case the "primary" member for the transaction is not available), and start the profile on that member. You would also change the location alias that maps to the one member, so that it maps instead to the other member (or you could map the alias to two members, and only start the alias on one member at any given time -- location aliases can be dynamically added, started, and stopped by way of the DB2 command -MODIFY DDF); so, no client-side changes would be needed to move single-threading for a transaction from one data sharing member to another.

A couple more use cases. What else can be accomplished via DB2 system profile monitoring? There are many possibilities. Consider this one: you have a certain DDF-using application for which you want to allow, say, 50 active threads. Easily done: if the application connects to DB2 using a particular authorization ID -- very commonly the case -- then set up a profile that is associated with the application's ID, and in the associated DSN_PROFILE_ATTRIBUTES row indicate that you want to MONITOR THREADS, that the threshold is 50, and the action is EXCEPTION. Note, then, that up to 50 requests associated with the application could be queued, if the 50 allotted DBATs are all in-use.

Or how about this: there are hundreds (maybe thousands) of people employed by your organization that can connect to one of your DB2 for z/OS systems directly from their laptop PCs. You know that a single individual could establish a large number of connections to the DB2 system, and you are concerned that, were that to happen, your system could hit its CONDBAT limit, to the detriment of other DDF users and applications (and maybe that's actually happened at your shop -- you wouldn't be the first to encounter this situation). How could you limit individuals' laptop PCs to, say, no more than 5 host connections apiece? Would you have to enter hundreds (or thousands) of rows in DSN_PROFILE_TABLE, each specifying a different user ID (or IP address or whatever)? That is what you WOULD have had to do, before a very important system profile monitoring enhancement was delivered with DB2 12 for z/OS (and retrofitted to DB2 11 via the fix for APAR PI70250). That enhancement, in a word: wildcarding. By leveraging this enhancement (explained below), you could limit EACH AND EVERY "laptop-direct" user to no more than 5 connections to the DB2 for z/OS subsystem by way of a single row in DSN_PROFILE_TABLE (and an associated MONITOR CONNECTIONS row in DSN_PROFILE_ATTRIBUTES).

More on wildcarding support for system profile monitoring: with DB2 12 (or DB2 11 with the fix for the aforementioned APAR applied), you can use an asterisk ('*') in the AUTHID or the PRDID column of DSN_PROFILE_TABLE (the latter can identify the type of client from which a request has come); so, an AUTHID value of 'PRD*' would apply to all authorization IDs beginning with the characters PRD (including 'PRD' by itself), and an asterisk by itself would apply to ALL authorization IDs (with regard to rows in DSN_PROFILE_TABLE, a DRDA request will map to the profile that matches it most specifically, so if there were a profile row for auth ID 'PROD123' and another row for auth ID '*', the former would apply to requests associated auth ID PROD123 because that is the more specific match).

You can also use wildcarding for the IP address in the LOCATION column of a row in SYSIBM.DSN_PROFILE_TABLE, but in a different form. For an IPv4 TCP/IP address, a wildcard-using entry would be of the form address/mm where mm is 8, 16, or 24. Those numbers refer to bits in the IP address. Here's what that means: think of an IPv4 address as being of the form A.B.C.D. Each of those four parts of the address consists of a string of 8 bits. If you want to wildcard an IPv4 address in the LOCATION column of a DSN_PROFILE_TABLE row, so that the row will apply to all addresses that start with A.B.C but have any possible value (1-254) for the D part of the address, the specification would look like this (if A, B, and C were 9, 30, and 222, respectively):

And note that a specification of applies to all IP addresses from which requests could come for the target DB2 for z/OS system. A similar convention is used for IPv6 addresses -- you can read about that in the text of the APAR for which I provided a link, above.

Why use this convention, instead of something like 9.30.222.* for addresses through, or an * by itself for all IP addresses? Because the convention used is already prevalent in the TCP/IP world, and in this case it made sense to go with the flow.

So, that's what I have to say about DB2 system profile monitoring. It's a great way to manage a DB2 for z/OS DDF workload in a more granular way than is offered by the ZPARMs CONDBAT, MAXDBAT, and IDTHTOIN (though those ZPARM values remain in effect in an overall sense when system profile monitoring is in effect). If you aren't yet using this functionality, think of what it could do for your system. If you are using system profile monitoring but haven't used the new wildcard support, consider how that enhancement could provide benefits for you. In an age of ever-growing DB2 DDF application volumes, system profile monitoring is a really good thing.