Back in 2018, I reviewed an organization's production Db2 for z/OS environment, and saw at that time the largest z/OS LPAR real storage size I'd ever seen: 1100 GB. The Db2 subsystem running in that LPAR had (not surprisingly) the largest buffer pool configuration I'd ever seen: 879 GB (referring to the aggregate size of all of the Db2 subsystem's buffer pools). Within that buffer pool configuration was (again, not surprising) the lagest single Db2 buffer pool I'd ever seen: approximately 262 GB (66,500,000 4K buffers).
Those "biggest I've ever seen" figures sustained that status until just recently, when I reviewed another production Db2 for z/OS system. The new "biggest I've ever seen" numbers are way larger than the previous maximums:
- A z/OS LPAR with 2.4 terabytes of real storage
- A Db2 subsystem with 1.7 terabytes of buffer pool resources.
- An individual buffer pool of approximately 560 GB in size (140,000,000 4K buffers)
Organizations that run production Db2 subsystems in z/OS LPARs that have large amounts of real storage are increasingly going with really large Db2 buffer pool configurations. That's a smart thing to do. Db2 allows for up to 16 TB of buffer pool space for a single subsystem, and going big is a definite plus for Db2 workload performance and CPU efficiency (just be sure that you don't over-burden a z/OS LPAR's real storage resource: you want to keep the LPAR's demand paging rate - available via an RMF Summary report - below 1 per second).
Here's the point I want to make with this blog entry: while really big buffer pools are great for Db2 application performance, don't focus solely on the size of a Db2 subsystem's buffer pool configuration. In addition to going big, take other steps to maximize the positive performance benefits of a large Db2 buffer pool configuration:
- Use large real storage page frames for the busier buffer pools - In my view, a buffer pool's GETPAGE rate (obtainable from a Db2 monitor-generated statistics long report) is the best indicator of a pool's busy-ness. My recommendation is to use large real storage page frames for every pool that has a GETPAGE rate (during a busy hour of the processing day) in excess of 1000 per second. Large real storage page frames enhance CPU efficiency for page access by reducing the CPU cost of translating virtual storage addresses to real storage addresses. Some related things to note:
- Page-fixing of buffers in a pool is a prerequisite for using large real storage page frames - This is done for a pool via a specification (in an -ALTER BUFFERPOOL command) of PGFIX(YES). Note that actually changing from not-page-fixed to page-fixed requires deallocation and reallocation of a pool - this usually happens as a consequence of recycling (i.e., stopping and then restarting) the associated Db2 subsystem. That deallocation and reallocation is also required to go from the default real storage frame size of 4K to a larger frame size.
- Use the appropriate real storage frame size - For a buffer pool that is at least 20 GB in size (a little over 5,000,000 buffers, if we're talking about a 4K pool), I recommend the use of 2 GB page frames (this is accomplished via a specification of FRAMESIZE(2G) in an -ALTER BUFFERPOOL command). For a pool that is smaller than 20 GB, I recommend the use of 1 MB page frames (FRAMESIZE(1M)). Note that if a pool is defined with PGSTEAL(NONE) (see below for information about such pools, which are known as "contiguous" pools), 2 GB frames cannot be used - you need to go with FRAMESIZE(1M) for a PGSTEAL(NONE) pool, unless objects assigned to the pool are quite small (e.g., less than 100 4K pages), in which case the default 4K frame size would be appropriate (for a PGSTEAL(NONE) pool, Db2 will use a given large real storage page frame for buffers that hold pages of one object assigned to the pool).
- Set the value of the z/OS parameter LFAREA appropriately - LFAREA, a parameter in the IEASYSnn member of the z/OS data set named SYS1.PARMLIB, specifies the amount of a z/OS LPAR's real storage resource that is to be managed in 1 MB (and, possibly, 2 GB) frames. I generally like to see an LFAREA specification that provides enough 1 MB (and 2 GB, when appropriate) frames to fully back FRAMESIZE(1M) (and, if relevant, FRAMESIZE(2G)) pools, but not much more than that. Why not much more? Because a lot of processes in a z/OS system can only use 4 KB page frames. If you think you might want to later enlarge some Db2 buffer pools that have a FRAMESIZE(1M) or a FRAMESIZE(2G) specification, you can make the LFAREA value for the frame size in question larger than you need at present - just don't go overboard with that.
- Use the output of the Db2 command -DISPLAY BUFFERPOOL(ACTIVE) DETAIL to verify that pools that have a preferred frame size of 1M or 2G are fully backed by frames of the desired size - If you see, in the output of that command, that a pool with a VPSIZE of 100,000 buffers and a preferred frame size of 1M has 70,000 buffers allocated in 1M frames and 30,000 buffers allocated in 4K frames, it means that there aren't enough 1 MB frames in the system to fully back the pool. Note that for a pool that has a preferred frame size of 2G, you might see some buffers allocated in 1M frames, even if the system has enough 2 GB frames to "fully" back the pool. Why might that be the case? Well, in the interest of not wasting a lot of space in a 2 GB page frame, Db2 won't use such a frame if it can't fill at least about 95% of the frame with buffers. Suppose you have a pool that is about 41 GB in size. Db2 will use twenty 2 GB frames for that pool, and the remaining 1 GB of buffers will be allocated in 1 MB frames, As far as I'm concerned, that is not at all a big deal. No problem.
- If you have any PGSTEAL(NONE) buffer pools, aim for zero activity in the overflow area of those pools - A PGSTEAL(NONE) buffer pool is also known as a "contiguous" buffer pool. Such a pool is intended to be used to completely cache in memory the objects assigned to the pool. A contiguous buffer pool can optimize efficiency for page access, but some of that advantage is lost if there is any activity in the pool's overflow area, which is where any buffer stealing would occur (if the pool were in fact not large enough to hold every page of every object assigned to the pool). Output of the Db2 command -DISPLAY BUFFERPOOL (ACTIVE) DETAIL will show if there is any activity in the overflow area of a PGSTEAL(NONE) pool. If you see that there is activity in the overflow area of a PGSTEAL(NONE) pool (unless it's a really small amount of activity), either take one or more objects out of the pool (by reassigning them to other pools) or make the pool larger (always keeping in mind that you want the demand paging rate of the z/OS LPAR to be less than 1 per second - don't over-burden the LPAR's real storage resource).
Hi Robert, we had problems in the past with PGSTEAL(NONE), when a REORG would take a long time to refresh the REORGed tablespace's pages.
ReplyDeleteCan you please tell us if this is a problem and what do we need to look out for?
Mare details on how Db2 handles Utilities against tablespaces in PGSTEAL(NONE) buffer pools?
OK, the thing to keep in mind is this: when a table space assigned to a PGSTEAL(NONE) buffer pool is online REORG-ed, the shadow data sets (which will become the new "original" data sets for the table space at the conclusion of the online REORG) are, of course, assigned to the same buffer pool as the table space being REORGed. If the table space is fairly large (or if the PGSTEAL(NONE) buffer pools is fairly small), that can put a lot of pressure on the PGSTEAL(NONE) pool, by potentially driving a lot of activity in the pool's overflow area. What to do about this? A few thoughts:
Delete* Assign the table space's indexes to a buffer pool other than the one to which the table space is assigned (because indexes are REORG-ed, too, when you online-REORG a table space).
* Make the PGSTEAL(NONE) pool larger, if the z/OS LPAR's real storage resource would permit that (as noted in the blog entry, you want the LPAR's demand paging rate to be less than 1 per second).
* If the table space in question is partitioned, do partition-level REORGs to reduce space requirements in the PGSTEA(NONE) buffer pool.
* Consider whether the table space in question even needs to be reorganized. At some sites, table spaces are periodically REORG-ed whether that is needed or not. REORG when needed, and don't when it's not needed.
* Try not to online REORG several objects assigned to the same PGSTEAL(NONE) pool at the same time.
* If none of the above are helpful or possible, you might need to consider reassigning the table space in question to a standard PGSTEAL(LRU) pool.
Robert
Hello Robert
ReplyDeleteI think I heard in a DB2 webcast, due to actual hardware capacity (cpu, disk controllers, real storage) , PGSTEAL(FIFO) can be better option than LRU because simplify BPOOL algorithm to keep it.
Can this be true?
In a bad scenario, if DBM1 takes a system dump (or IBM asked for one) these big BPOOL will freeze lpar for several seconds, correct?
(of course we don´t expect this to happen...)
Thanks!
I'd say that in the very large majority of situations, the CPU cost associated with tracking references to pages in a buffer pool when PGSTEAL(LRU) is in effect is way more than offset by the CPU savings related to the reduced read I/O activity typically delivered by the LRU buffer steal algorithm. I see PGSTEAL(FIFO) as very much a niche option - beneficial only in unusual situations. PGSTEAL(FIFO) MIGHT be beneficial for.a buffer pool when pages are rarely re-referenced after initially being read into memory - like I said, an unusual situation.
DeleteAs for a DBM1 dump, the impact of such an occurrence on a z/OS LPAR will depend on a number of factors, including the LPAR's real storage resource and the number of engines (processors) configured for the LPAR. I think that in most cases the impact of a DBM1 dump will be secondary to the impact of the Db2 failure situation that triggered the dump.
Robert
Thank you very much Robert
DeleteHello Robert
ReplyDeleteCan this increase in bufferpools size affect db2 commit (gets heavier)?
Do we need to review CHKFREQ in ZPARM?
I don't see that larger buffer pools would make commit processing more expensive, nor do I see that larger buffer pools would impact the preferred value of CHKFREQ in ZPARM. With regard to CHKFREQ, my preference is to go with time-based Db2 checkpoint frequency, and I like the default specification of 3 minutes between checkpoints.
DeleteWhat larger buffer pools could impact is the preferred value of the VDWQT specification for a given pool. A larger buffer pool might cause VDWQT to be hit less frequently, and that could lead to more synchronous write actions for the buffer pool than you'd like. As a general rule, I like to go with the default VDWQT value of 5 for a buffer pool. I might want to adjust that value downwards if I saw (via my Db2 monitor or in the output of the Db2 command -DISPLAY BUFFERPOOL(ACTIVE) DETAIL) that synchronous writes were outnumbering asynchronous writes for a given pool. An exception to that rule: I wouldn't worry about synchronous writes for a pool outnumbering asynchronous writes if the volume of write activity were quite low for the pool (e.g., in the single digits per second).
Robert
Thank's Robert!
DeleteLFAREA for 1M frames is no longer a pre-allocation since z/OS 2.3 - it's now a limit. So you can set this higher than your current allocations safely without affecting the availability of 4K pages - or worry about pressure on 4K pages causing 1M pages to be broken down.
ReplyDeleteThat's very useful information - thanks for posting this comment.
DeleteRobert