I had an interesting exchange recently with a Db2 for z/OS systems programmer. This individual had been asked to assist in determining the cause of a performance problem impacting a Db2-accessing application. The sysprog shared with me a Db2 monitor-generated accounting report showing activity for the application process, and a Db2 monitor statistics report covering the same time period for the same Db2 subsystem. In those two reports were the clues that pointed to the source of the application performance problem, and with the probable problem source identified the associated mitigating actions were readily determined. In this blog entry, I'll take you through my analysis of the Db2 monitor-provided accounting and statistics information, and the rationale behind my recommended steps for resolving the problem. My aim is not only to shed light on a particular performance-affecting issue and related remediating moves, but also to illustrate a methodical approach for analyzing Db2 application performance issues in general.
Input for analysis
I find Db2 monitor generated accounting and statistics reports to be extremely useful for analysis of application performance problems. In both cases, what you want is the detailed form of the report. In the case of IBM's OMEGAMON for Db2 performance monitor (the one with which I'm most familiar), you're talking about the ACCOUNTING REPORT - LONG and the STATISTICS REPORT - LONG (for other Db2 monitor products, these reports might have titles like, "detailed summary of accounting information" or "statistics detail report"). For the accounting report, your preference is to see activity pertaining exclusively to the application process for which the performance issue has arisen. This is usually done by using the monitor's data-filtering capabilities to include, for report-generation purposes, only the Db2 accounting records of interest (an accounting report is basically information from Db2 accounting trace records, formatted for readability). Db2 accounting trace records have all kinds of identifier fields, so you can get pretty specific. In the case about which I'm writing here, the relevant filtering criteria were the authorization ID of the application process of interest, and the FROM and TO times that bracketed the period during which the performance problem occurred.
As for the statistics report, what you want is one that covers the same time period as the accounting report (same FROM and TO times), for the same Db2 subsystem.
Looking for clues
What the Db2 sysprog had been told by the development team is that the application in question started out performing well, and then slowed down a lot (important input). The sysprog and I looked first at the accounting report, and in doing that we focused initially on the "class 3 suspension" information (this information, obtained from Db2 accounting trace class 3 records, has to do with "known" wait events, as opposed to "not accounted for" time, about which I'll comment momentarily). Why did we look there first? Because (in my experience), when a Db2-accessing process starts out performing well and then slows way down, it's often due to a substantial increase in one or more of the "wait times" captured by accounting trace class 3 (yes, an application slowdown could be related to a significant CPU time increase, but I've seen that less often than I've seen large wait time increases).
Looking at the class 3 times for the application process, what jumped out was a very large value for the average DB2 LATCH time ("average" is average per accounting trace record, which typically equates to average per transaction or average per batch job, depending on the workload type). Here, I mean "large" in terms of average DB2 LATCH time being a large percentage of average TOTAL CLASS 3 time. Usually, DB2 LATCH time is a very small percentage of TOTAL CLASS 3 time, with "wait time" categories such as SYNCHRON DATABASE I/O and OTHER READ I/O accounting for the bulk of TOTAL CLASS 3 time. A (proportionately) really large DB2 LATCH time is usually an indicator that something's not right.
The first thing I look at when I see unusually large DB2 LATCH time for a Db2-accessing process is the "in-Db2 not-accounted-for time" for that process. Several Db2 monitor products calculate that for you - in an IBM OMEGAMON for Db2 accounting report, the field is labeled NOTACC, and it's shown, on a sideways bar chart at the top of an accounting report, as a percentage of average in-Db2 elapsed time. If you need to calculate this figure for yourself, the denominator is average in-Db2 elapsed time (aka "class 2" elapsed time), and the numerator is average in-Db2 elapsed time minus in-Db2 CPU time (general-purpose plus zIIP, or "specialty engine," CPU time) minus TOTAL CLASS 3 time. In other words, it's the percentage of in-Db2 elapsed time that is not CPU time and not "identifiable" wait time. For a transactional application process (as was the case for the situation about which I'm writing), as a general rule you want in-Db2 not-accounted-for time to be less than 10%. If that figure is substantially greater than 10%, it's indicative of a CPU-constrained environment, and if the environment is highly CPU-constrained then DB2 LATCH time can get really large (as pointed out in an entry I posted to this blog a few years ago).
Well, in this particular case the average in-Db2 not-accounted-for time for the application process was 11% of in-Db2 elapsed time - a little on the high side for a transactional process, but not high enough to explain a really large DB2 LATCH time. With that cause of elevated DB2 LATCH time pretty much eliminated, it was time to turn to the breakdown of latch wait events for different latch categories, and that's where the statistics report comes in. In a Db2 monitor-generated statistics long report, the latch suspend count information will likely look something like this (what you see below is from an OMEGAMON for Db2 statistics long report, but it is NOT from the report I reviewed with the Db2 systems programmer - we jointly viewed that report in a web meeting, and I do not have a copy of the report):
--------- -------- -------- -------- --------
LC01-LC04 0.00 0.00 0.00 0.00
LC05-LC08 0.00 0.74 0.00 0.41
LC09-LC12 0.00 0.02 0.00 0.32
LC13-LC16 0.00 12.89 0.00 0.00
LC17-LC20 0.00 0.00 0.01 0.00
LC21-LC24 0.04 0.00 1.96 2.84
LC25-LC28 0.12 0.02 0.01 0.00
LC29-LC32 0.06 0.04 0.00 0.28
What I saw in the report I reviewed in the web meeting with the Db2 sysprog (and again, that's NOT what you see above - the snippet above is provided so that you can see what the latch suspend count information looks like in a statistics report) was a particularly high value for latch class 6 suspend events (that would be in the position highlighted in green in the sample report snippet above). What is latch class 6? It has to do with index page split activity in a Db2 data sharing environment (by the way, a handy page for seeing the activities associated with various Db2 latch classes is this one from the OMEGAMON for Db2 documentation).
Let's unpack that. An index page split occurs when Db2 has to insert an entry in an index page because of an insert (or an update of an indexed column) and that page is full. In that situation, a portion of the entries in the page will be moved to what had been an empty page in the index, so that there will be room in the formerly-full page for the new entry. What does Db2 data sharing have to do with this (and in the environment about which I'm writing, Db2 is running in data sharing mode)? In a data sharing system (versus a standalone Db2 subsystem), an index page split action has a greater impact on throughput because it forces a log-write operation.
Seeing the high level of index page split activity suggested by the numerous latch class 6 wait events, we turned again to the accounting report to see the average number of insert operations executed by the performance-impacted application. Sure enough, we saw that this was an insert-intensive process - more by the average number of rows inserted, versus the number of INSERT statements executed (the average number of rows inserted, per the accounting report, was about 100 times larger than the average number of INSERT statements executed, indicating use of block-level inserts by the application).
The elevated count of latch class 6 suspend events (related to index page splits) and the insert-intensive nature of the process also dove-tailed with the observed "started out fine, then slowed down" behavior of the application: in all probability, when the process started there was a pretty good amount of free space in leaf pages of indexes on the table(s) into which rows were being inserted. After a while these "holes" in index leaf pages filled up, and that resulted in a large number of index page split actions to make space for new index entries, and THAT - partly due to the fact that this was a Db2 data sharing system - had a majorly negative impact on application performance (the keys of affected indexes were clearly not of the continuously-ascending variety, because index page split actions are not required for an index defined on a continuously-ascending key).
With the source of the performance problem identified, the next matter to consider was...
What to do about it?
The Db2 sysprog and I discussed two problem-mitigating actions - one an "absolutely do" and the other a "maybe do." The "absolutely do" step was to increase the amount of free space in indexes to accommodate new entries. That step, in turn, was comprised of two sub-steps, each of which is applicable to indexes defined on non-continuously-ascending keys. Sub-step one: increase the index's PCTFREE value. Whereas the default PCTFREE value for an index is 10, a value of 20 or 25 might make more sense for an index defined on a non-continuously-ascending key for an insert-heavy table. Sub-step two: increase the index's FREEPAGE value. The default FREEPAGE value is 0. Here's why boosting the FREEPAGE value - for example, to 5 (in which case there would be an empty index page after every 5 pages containing index entries) - can be helpful for an index defined on a non-continuously-ascending key for an insert-heavy table: as previously mentioned, when an index page is split a portion of that page's entries are moved to a previously empty page in the index. If FREEPAGE 0 (the default) is in effect, the only empty index pages will be at the very end of the index - potentially a long way from the page that was split. That situation creates a drag on performance through degradation of the index's organization (reflected in the LEAFDIST value for the index - or index partition, in the case of a partitioned index - in the SYSIBM.SYSINDEXPART catalog table). With a non-zero and relatively low value for FREEPAGE (meaning, an empty page following each relatively-low-number of populated pages), when an index split does occur then there should be a "nearby" empty page into which entries from the full page can be moved.
Both PCTFREE and FREEPAGE can be changed for an index via an ALTER INDEX statement, and both take effect when the index is subsequently reorganized (or loaded). The larger PCTFREE value will reduce index page split activity between REORGs, and the non-zero FREEPAGE value will reduce the impact of page splits if they do occur.
And what about the "maybe do" step? That would be an increase in the size of the index's pages, from the default of 4 KB to maybe the maximum of 32 KB. How can that help? Here's how: because a 32 KB index page (for example) can hold 8 times as many entries as a 4 KB page, going to 32 KB-sized pages for an index (via an ALTER INDEX statement that assigns the index to a 32K buffer pool) can potentially result in an 87.5% reduction (seven eighths) in page split activity for an index, other things being equal (e.g., same rate of inserts for the underlying table). Why is this a "maybe do" thing versus an "absolutely do" thing? Because if access to table rows through the index (e.g., for queries) is truly random in nature with respect to key values, 32 KB-sized pages could mean somewhat less-effective use of buffer pool resources versus 4 KB-sized pages. It's a matter, then, of what's more important in a particular situation: is it minimizing index page split activity, or maximizing the effectiveness of buffer pool resources for random patterns of row access by applications?
I hope that this entry has provided you with some "news you can use." Utilize Db2 monitor accounting and statistics reports to determine the source of an application performance problem, then take appropriate remedial action.
One more thing: Db2 13 for z/OS made it a lot easier to verify that index page splits are an issue, through enhanced instrumentation (new IFCID 396, associated with statistics trace class 3, which is on by default, indicates when an index page split operation takes more than 1 second, which would be unusually long) and through the new REORGTOTALSPLITS, REORGSPLITTIME and REORGEXCSPLITS columns of the SYSIBM.SYSINDEXSPACESTATS real-time statistics table in the Db2 catalog.