Monday, April 2, 2018

Another Db2 for z/OS Blog You Should Check Out

If you're reading this, it's probably because you've found blogs to be a useful source of technical Db2 for z/OS information. I want to direct your attention to one of the newer kids on the Db2 blog block: Db2 for z/OS News from the Lab.

What makes the Db2 for z/OS News from the Lab blog particularly noteworthy? It's written by people who are part of IBM Db2 for z/OS development organization - people who call the IBM Silicon Valley Lab home base. These folks, needless to say, know Db2 for z/OS very well. Additionally, these people interact with Db2 for z/OS users all the time, and those interactions provide inspiration for blog entries that can clarify and highlight various facets of the Db2 whole.

So, get wise and get the inside scoop by visiting the Db2 for z/OS News from the Lab blog. Even better, become part of the blog's community - information on doing that can be found in one of the blog's entries. Read and learn!

Wednesday, March 28, 2018

Db2 for z/OS: CACHEPAC is Another ZPARM You Should Check

Sometimes, the default value of a Db2 ZPARM parameter is changed to reflect new realities of Db2 workloads and of systems on which Db2 for z/OS runs. When that happens, a Db2-using organization may stick with the old default value for the ZPARM, and going that route can have a negative impact on the performance of the Db2 system. Case in point: CACHEPAC, the subject of this blog entry.

I've blogged before about ZPARMs that are regularly set to sub-optimal values. I've also posted entries (part 1, part 2, and part 3) on boosting Db2's use of memory so as to improve CPU efficiency for a Db2 workload. CACHEPAC is related to both of those topics.

What is the purpose if CACHEPAC? It specifies the amount of virtual storage, in the Db2 database services address space (aka DBM1), that will be used to cache package authorization information (like, what authorization IDs have the EXECUTE privilege on package XYZ). Is caching that package authorization information important? Yes. Why? Volume - as in, volume of requests for package execution.

Get yourself a statistics report - long generated by your Db2 monitor (depending on the monitor product used, that might be called a statistics detail report), and check the section of the report that contains figures for EDM pool-related activity (alternatively, look at an online display of EDM pool activity provided by your Db2 monitor). Look for a field with a label like PT REQUESTS. PT is short for package table, and this field shows the frequency of requests to execute packages (or sections of packages). For a production Db2 subsystem, you might see hundreds, or even thousands, of requests to execute package sections per second. Having to check the SYSIBM.SYSPACKAUTH table in the catalog for each and every one of those package execution requests would be a drag on performance, for sure; thus, the in-memory package authorization cache specified via CACHEPAC in ZPARM.

It was with Db2 10 for z/OS that the default value for CACHEPAC changed, and changed pretty dramatically. In a Db2 9 environment, the default value for CACHEPAC was 100 KB. With Db2 10, the CACHEPAC default value changed to 5 MB - 50 times larger than before. Why? I suspect the change was made by the Db2 for z/OS development team for several reasons. For one thing, Db2 systems have ever-larger numbers of users and applications (that's more authorization IDs needing package-related privileges), and ever-larger numbers of packages (further upping space needs for package authorization information). Additionally, z/OS LPAR memory resources are getting bigger all the time (hundreds of GB of memory for one z/OS LPAR is not unusual these days), and with more available memory "real estate," why not use more of it for package authorization information caching, to enhance system performance?

Along with Db2 10's boost of the CACHEPAC default value, the maximum value for the parameter was increased - to 10 MB, from 5 MB.

OK, so it's not uncommon to see a Db2-using organization stick with the old default value for a ZPARM when a new default comes along. This might be due, in some cases, to inertia as much as anything. Trouble is, a 100 KB package authorization cache may prove to be WAY smaller than it should be for a Db2 workload that's large by today's standards. I saw a situation of that nature quite recently, in looking over a Db2 monitor statistics detail report sent to me by the Db2 team at a company with which I was working. The report showed that the package authorization cache for a Db2 system that was having occasional performance problems was severely undersized. How could I see that? Easy. In such a report, find a section with a heading like AUTHORIZATION MANAGEMENT. In that section of the report, find the field labeled something like AUTH UNSUCC-PKG-CACHE. That is where you'll see how often a package authorization check could NOT be successfully performed using the package authorization cache. What would you LIKE to see here? Maybe a few package authorization checks per second (in the single digits, or less, per second) that could not be serviced using information in the package authorization cache. What did I see in that report I just referenced? 1098 package authorization requests per second that were not successfully performed via the package authorization cache. Why was this so? It was so because, on a busy Db2 system doing a lot of work for applications, the package authorization cache size, per CACHEPAC in ZPARM, was too small - probably way too small.

So, here's what you should do: use your Db2 monitor to generate a statistics report - long (or, if you can't do that, check an online display of subsystem statistics provided by the Db2 monitor), and find in that report the AUTHORIZATION MANAGEMENT section. In that section, find the field with the label like PKG-AUTH UNSUCC-CACHE. What's the frequency of package authorization checks NOT accomplished by way of the package authorization cache? If that number is in the low single digits (or less) per second, good. If it's up in the double digits per second or higher, the value of CACHEPAC in ZPARM should be increased. If the frequency of package authorization requests not successfully performed using the package authorization cache is really high, like north of 100 per second, CACHEPAC should probably be substantially increased.

What is the value of CACHEPAC for your production Db2 subsystems? If it is less than the default (since Db2 10) of 5 MB, it should probably be increased to 5 MB. If it is at 5 MB and you still see a lot of package authorization requests NOT successfully performed using the package authorization cache, CACHEPAC should probably have a value greater than 5 MB (remember, that value can go up to 10 MB).

Do you have the memory resources on your z/OS LPAR to support an increase in the CACHEPAC value? Probably (if you can't use a few more MB of memory for Db2 without putting the z/OS LPAR's real storage resource under an undesirably high level of pressure, the LPAR likely has a too-small memory allocation). How can you tell if the LPAR's memory is stressed? Check (or have a z/OS systems programmer check) the LPAR's demand paging rate in an RMF (or equivalent z/OS monitor) CPU Summary Report. If the LPAR's demand paging rate is zero, memory is not overly stressed and you should be able to use more for the Db2 package authorization cache (if a CACHEPAC size increase is seen as being needed). I'd also be comfortable with a CACHEPAC size increase if the associated LPAR's demand paging rate were a very small non-zero value (i.e., less than 1 per second). If the LPAR's demand paging rate is 2-3 per second or higher, I'd like to see more memory allocated to the LPAR before substantially increasing CACHEPAC (or, if more memory can't be provided for the LPAR, consider reducing some of Db2's memory use until the demand paging rate is less than 1 per second, and THEN make CACHEPAC larger if it is too small and failing to satisfy a lot of package authorization requests per second).

And that's that. When most all package authorization checks can be successfully accomplished using information in the Db2 package authorization cache, it's a good thing. If you're not there for one or more of your production Db2 subsystems, try to get there.

Monday, February 26, 2018

Db2 for z/OS: DDF, zIIP Engines, and SMT2

SQL statements executed by way of the Db2 for z/OS distributed data facility (DDF) are not, of course, the only workload that uses zIIP MIPs, but in my experience this does tend to be the main driver of zIIP utilization in a z/OS system on which Db2 runs. I posted an entry to this blog, a few years ago, on the importance of avoiding zIIP engine contention. In that blog entry, I described how to calculate the "zIIP spill-over ratio" for a Db2 DDF workload (the percentage of zIIP-eligible work that ends up being processed by general-purpose engines), and pointed out that this ratio should be less than 5% for optimal performance and CPU cost-efficiency (less than 1% is outstanding). That rule-of-thumb begs the question, "How 'hot' can I run my zIIP engine(s) while avoiding a higher-than-desired zIIP spill-over ratio?" The answer to that question, as pointed out in the blog entry, depends on the number of zIIP engines available to the z/OS LPAR in which the Db2 subsystem of interest is running: the more zIIP engines an LPAR has, the higher the rate at which they can be utilized without overly-high zIIP spill-over becoming an issue. In the entry I'm writing now, I want to supplement the blog entry just referenced (and pointed to by the preceding hyperlink) with information pertaining to an IBM Z server processor technology called simultaneous multithreading, or SMT (or SMT2, as explained below).

SMT2, introduced with the IBM z13 line of mainframe servers and available as well with the newer z14, refers to a technology whereby two processes (thus the 2 in SMT2) can be active at the same time on one "engine" (i.e., one core). You can find more information about SMT2 in an entry on the z13 that I posted to this blog a few years ago.

One thing that SMT2 affects is the look of an IBM RMF CPU Activity Report (that report, by the way, is very useful when it comes to understanding the z/OS environment in which a Db2 subsystem operates). One section of an RMF CPU Activity Report shows the average utilization (usually over a 15-minute interval of time) of the engines available to a z/OS LPAR. For a system running on an IBM Z server with SMT2 enabled, that section of the report might look something like this (with color highlighting added by me):


---CPU---    ---------------- TIME % ----------------
NUM  TYPE    ONLINE    LPAR BUSY    MVS BUSY   PARKED
 0    CP     100.00     4.28         4.15        0.00
 1    CP     100.00     1.41         1.42        0.00
TOTAL/AVERAGE           2.84         2.79           
 2    IIP    100.00    76.08        75.23        0.00
                                    71.13        0.00
 6    IIP    100.00    67.62        66.22        0.00
                                    64.94        0.00

TOTAL/AVERAGE          71.85        69.38

Here are some things to notice about the above report snippet: first, see those numbers in the "MVS BUSY" column that I highlighted in red? See how there are two such numbers for each physical zIIP engine (labeled as type "IIP")? Know what that means? It means that SMT2 is active for those zIIP engines. Could you see two MVS BUSY numbers for one general-purpose engine (labeled as type "CP")? No. Why? Because SMT2 can be activated for zIIP engines and also for IFL engines (engines dedicated to Linux systems running on an IBM Z server), but not for general-purpose engines (more on "can be activated" in a moment).

A second thing that can look different in an RMF CPU Activity Report, when SMT2 is active for a system's zIIP engines, is the section, just below the above-referenced CPU utilization section, that shows information about "in-and-ready" tasks and how often these tasks had to wait for dispatch because all of the LPAR's engines were busy. That section of the report might look like this (again, color highlighting was added by me -- and this snippet and the CPU utilization snippet shown above are from the same RMF report):


-----------------------DISTRIBUTION OF IN-READY WORK UNIT QUEUE
 NUMBER OF              0    10   20   30   40   50   60   70
 WORK UNITS     (%)     |....|....|....|....|....|....|....|...
                                            
<=  N          40.3     >>>>>>>>>>>>>>>>>>>>>
 =  N +   1     5.5     >>>
 =  N +   2    52.3     >>>>>>>>>>>>>>>>>>>>>>>>>>>
 =  N +   3     0.9     >
<=  N +   5     0.4     >
<=  N +  10     0.0
<=  N +  15     0.0
<=  N +  20     0.0
<=  N +  30     0.1     >
<=  N +  40     0.0
<=  N +  60     0.0
<=  N +  80     0.0
<=  N + 100     0.0
<=  N + 120     0.0
<=  N + 150     0.0
>   N + 150     0.0

N = NUMBER OF PROCESSORS ONLINE (6.0 ON AVG)

See that line, highlighted in red, that tells you the value of "N"? Know what "N" is? It's the number of engines available to the LPAR for which the report was run (if you see a non-integer value for N, such as 8.5, it means that the LPAR has part of an engine, or parts of several engines, available to it -- that's indicated by non-zero values in the PARKED column of the CPU utilization section of the report referenced previously). Note that in this case, N = 6, even though the LPAR has, as we saw earlier, four physical engines -- two general-purpose engines and two zIIP engines. N = 6 because two tasks can be dispatched to each of the two zIIP engines at one time (so, a total of 6 tasks can be executing at one time: two on the two general-purpose engines and four on two the zIIP engines, each of which can handle two tasks simultaneously).

Know what else can look different with SMT2 activated for your zIIPs, versus SMT2 not being activated for those engines? The zIIP spill-over ratio. Here's what I mean by that: suppose you had a Db2 for z/OS subsystem, with a significant DDF workload, running on a z13 or z14 mainframe for which SMT2 had not been activated for the zIIP engines. Suppose that on that z/OS system, with X number of zIIP engines, the zIIP spill-over ratio was Y. If you subsequently activated SMT2 for the zIIP engines, what would be the effect on the zIIP spill-over ratio for the Db2 DDF workload? I think it's pretty safe to say that the zIIP spill-over ratio would be less than Y (i.e., less than what it had been with SMT2 not activated for the zIIP engines).

How substantial is the effect of activating SMT2 for zIIP engines? Is it like turning two zIIP engines into four zIIP engines? No. Why? Because the speed at which a given task is executed is somewhat reduced when that task is dispatched to an engine for which SMT2 has been activated, versus the same engine with SMT2 not activated. Generally speaking, when SMT2 has been activated for a zIIP engine, that engine is capable of processing about 40% more work than before; so, enabling SMT2 for two zIIP engines is roughly akin to turning them into 2 X 1.4 = 2.8 engines, capacity-wise.

Now, I keep referring to SMT2 "being activated" for a zIIP engine. That tells you that a zIIP (on a z13 or z14 server) can run in either "uni-thread" or SMT2 mode. Is there any reason NOT to activate SMT2 for a system's zIIP engines? About the only scenario possibly favoring "uni-thread" mode for zIIPs that comes to my mind is one involving a zIIP-eligible process that is single-threaded and batch-like in nature, and which requires the fastest possible single-thread zIIP performance. For a workload that is primarily transactional in nature (true for most DDF workloads I've seen), I would lean strongly towards activating SMT2 for the zIIP engines of the associated z/OS LPAR. The expected result should be greater throughput and less in the way of zIIP-eligible work spilling over to general-purpose engines.

Sunday, January 21, 2018

Db2 12 for z/OS and Insert Algorithm 2: Faster, More-Efficient INSERTs than Ever Before

Near the end of the blog entry I most recently posted prior to this one, I mentioned that a specification of MEMBER CLUSTER for a Db2 for z/OS universal table space could help maximize INSERT throughput and efficiency, in situations where the importance of that objective outweighs that of keeping data rows physically ordered by the target table's clustering key. I also noted that MEMBER CLUSTER sets you up to take advantage of a Db2 12 for z/OS enhancement called (rather cryptically) insert algorithm 2.

So, what's insert algorithm 2? I'll tell you, but before getting to the "what" of this Db2 12 feature, I want to spend a little time on the "why."

I recall seeing, way back in the mid 1980s, not long after the introduction of what was then called Db2 for MVS (the fruit of an IBM research effort that led to the invention of relational database technology), an IBM "red book" about the software product. This document cautioned users against thinking of Db2 as a database management system suitable for high-volume batch and online transactional workloads: Db2 Version 1 Release 1 was intended to support business intelligence-type queries, and was not designed to provide the performance organizations generally required for operational, "run the business" applications.

Well, that advice was widely ignored in the user community. Relational technology and the SQL data-interface language (also invented by IBM) delivered such an advance in programmer productivity versus traditional data stores that companies jumped on Db2 as a foundation for operational applications, in spite of the initial performance challenges. IBM responded to this development by providing, in successive versions and releases of Db2, features and functions that vastly boosted the ability of the DBMS to, among other things, ingest large amounts of new data records in compressed time frames. What were the bottlenecks holding back INSERT throughput and efficiency? Log processing? Index maintenance? Lock contention? Latch contention? Instruction path length? Client-server network management? Smart people at the IBM Silicon Valley Lab kept flattening these INSERT speed bumps, until organizations were routinely able to drive thousands of inserts per second into Db2 tables. And still, in some cases, that wasn't fast enough,

So we come to the Db2 11 for z/OS time frame, and the question is, how do you take Db2 INSERT throughput from really high to super-duper high? You (if you're in the Db2 development organization) do what you've always done: you identify what's holding throughput back in the current technology environment. It was determined that in a Db2 11 system, a key INSERT bottleneck involved the search for space, in pages of a target table, to hold incoming rows. How do you address that? Here, a page was taken from the Db2 playbook of long ago: asynchronous, anticipatory processing is the ticket to success.

Long ago, read I/O was seen as a major drag on Db2 performance, were it to be always handled in the traditional, synchronous way (i.e., read a single page from disk into memory on-demand, at the time a program needed the contents of the page). The solution was something that, today, we can scarcely imagine Db2 being without: prefetch. Through prefetch, Db2 reads batches of table space and/or index pages into memory, in anticipation that the pages will be requested by a process that seems likely to scan a significant number of pages. Because prefetch reads are anticipatory in nature, application elapsed time is reduced: when a prefetch-driving application process requests a page, it is likely that the page has already been read into a Db2 buffer pool in memory, so there is no need to suspend the application's task in order to perform a single-page, on-demand read. Because prefetch reads are asynchronous in nature, application elapsed time is further reduced: prefetch reads are executed by way of Db2 tasks (specifically, by preemptable SRBs in the DB2 database services address space), so application tasks can remain "on task" with regard to processing data (versus having to bring data into memory).

Apply those same concepts - anticipatory, and asynchronous - to an insert-intensive application process, and - presto! - you have a good understanding of the essential nature of insert algorithm 2.

The "how" of enabling insert algorithm 2 is largely table space-based - it can be used only for inserts into universal table spaces that are defined with the MEMBER CLUSTER option (that option can be specified in a CREATE TABLESPACE or an ALTER TABLESPACE statement). What is additionally required is telling Db2 that insert algorithm 2 is to be used for universal, MEMBER CLUSTER table spaces. That can be done at a Db2 subsystem level (at function level V12R1M500 or above) by way of a new (with Db2 12) ZPARM parameter, called DEFAULT_INSERT_ALGORITHM. When the value of that parameter is set to 2 (the default value), insert algorithm 2 will be automatically used for universal table spaces defined with MEMBER CLUSTER (unless that behavior is overridden at the table space level - more on that to come). When DEFAULT_INSERT_ALGORITHM is set to 1, insert processing will be done in the traditional way for universal MEMBER CLUSTER table spaces, unless insert algorithm 2 has been explicitly enabled for a given table space (again, table space-level control over insert algorithm 2 will be covered momentarily). When DEFAULT_INSERT_ALGORITHM is set to 0, insert algorithm 2 functionality is disabled for the subsystem - it cannot be used for any table space.

OK, table space-level control: if the ZPARM parameter DEFAULT_INSERT_ALGORITHM is set to 2 (meaning, insert algorithm 2 is used by default for universal MEMBER CLUSTER table spaces), you can de-activate the functionality (i.e., go with traditional insert processing) for an individual table space by specifying INSERT ALGORITHM 1 for the table space (either at CREATE TABLESPACE time, or via ALTER for an existing MEMBER CLUSTER universal table space). Similarly, if the ZPARM is set to 1 (traditional insert processing will be in effect, by default, for MEMBER CLUSTER universal table spaces), and you want to override that value for a particular MEMBER CLUSTER universal table space (i.e., you want insert algorithm 2 to be in effect for that table space), you can specify INSERT ALGORITHM 2 for the table space (again, either in a CREATE TABLESPACE or an ALTER TABLESPACE statement). Note that a value of 0 has a different meaning for the table space specification versus the ZPARM. For the DEFAULT_INSERT_ALGORITHM parameter in ZPARM, a value of 0 means that insert algorithm 2 functionality is disabled for the Db2 subsystem. For a MEMBER CLUSTER universal table space, INSERT ALGORITHM 0 (the default value for CREATE TABLESPACE, and a specification that is valid only for CREATE TABLESPACE) means that the insert algorithm that will be in effect for the table space is the one specified via DEFAULT_INSERT_ALGORITHM in ZPARM.

Let's say that insert algorithm 2 is in effect for a MEMBER CLUSTER universal table space, and an insert process targeting the table in that table space gets going. What happens? Here's what happens: Db2 provides a "pipe" (an in-memory area) that will be used to supply the insert process (or processes - several can use one pipe) with pages, belonging to the partition receiving new rows, that have space to hold new rows (if several partitions of a table space are receiving new rows, each partition will have its own fast-insert pipe). If row-inserting processes are taking pages from the pipe, what's putting pages into the pipe? That would be a Db2 task that puts insert-ready pages into the pipe, in an asynchronous manner (that is, in a way that does not interrupt the tasks of inserting processes, as a Db2 prefetch task does not interfere with the tasks of page-accessing programs).

What this auto-filled page pipe approach does is remove, for an inserting process, the need to find pages of a partition that have space enough to hold new rows. Even though the space search mechanism used for traditional insert processing is quite efficient (especially for segmented table spaces, thanks to the relatively precise "space available" information maintained in space map pages - and all universal table spaces are segmented), insert throughput can be expected to increase when an inserting process can simply pull "insert space available" pages from a fast-insert pipe versus having to locate such pages itself.

How much will insert throughput increase when insert algorithm 2 versus algorithm 1 (traditional insert processing) is used, all other things being equal? You probably know the answer to that question: it depends. In some cases, insert throughput might increase by a few percentage points. In other cases, the throughput improvement could be substantially greater (some tests have shown a throughput increase in the vicinity of 25%, and even larger increases have been observed in other test scenarios). CPU efficiency gains can also be quite varied - in some cases, in-Db2 CPU time (aka "class 2" CPU time) may decrease only slightly for an inserting process, while in other cases in-Db2 CPU efficiency might improve by 10%, or closer to 25%. Why the variance in throughput and CPU savings for inserting processes? One factor is indexes: generally speaking, the fewer the number of indexes defined on a target table, the greater the performance improvement delivered by insert algorithm 2 will be (the feature is used for inserts of rows into table pages, not for the insert of entries into index pages). The other main factor is the presence of bottlenecks other than row-accommodating-page-search that might "bubble to the surface" and constrain the performance of an insert process, when insert algorithm 2 is in effect. For example, consider a case in which inserts are coming from network-attached processes (i.e., DDF-using applications) that are running on servers that are physically separate from the Db2 subsystem, versus the same processes running in Linux systems on the IBM Z server that also houses the Db2 for z/OS subsystem. In the latter case, the greater network throughput that can be available thanks to IBM HiperSocket technology (memory-to-memory communications links) might result in a comparatively larger performance improvement with insert algorithm 2 in effect, because a more-pronounced network-related bottleneck might constrain the fast-insert improvement achieved with insert processes running on "boxes" that are physically separated from the Db2 server (to put this another way: when insert algorithm 2 removes one insert performance bottleneck, another bottleneck could come into play).

So, your mileage may vary, as they say.

How about the memory impact of insert algorithm 2? Those page-supplying pipes occupy some space, don't they? Yes, but a fast-insert pipe should take up less space than a data-compression dictionary (remember that every partition of every compressed table space has an associated compression dictionary), and at many sites there will be quite a few more COMPRESS YES table spaces than table spaces for which insert algorithm 2 is in effect (recall that insert algorithm 2 is only applicable to universal table spaces defined with MEMBER CLUSTER). Translation: if the real storage resource of a z/OS system were under a good bit of pressure (i.e., if the system's demand paging rate were in the mid-single-digits per second or higher), I might want to add memory to that system before making extensive use of insert algorithm 2; otherwise, you are likely good to go from a memory perspective.

IBM keeps boosting the performance of Z servers. Taking full advantage of these hardware enhancements depends in large part on software design, and Db2 12's insert algorithm 2 is part of that story. With the new insert algorithm in effect, insert rates of hundreds of thousands of rows per second have been observed in tests of Db2 12 systems (in some tests involving tables on which no indexes are defined, insert rates exceeding 1 million per second have been achieved). Does that kind of throughput require a substantial amount of CPU processing capacity? Yes. The insert-algorithm 2 value proposition is this: if you have Z horsepower to apply to an insert throughput challenge, Db2 12 lets you put it to work like never before.