Last month, IBM announced the z13 -- the latest generation of the mainframe computer. The z13 is chock full of great technical features, and there are already plenty of presentations, white papers, and "redbooks" in which you can find a large amount of related information. Going through all that information is an option, but you might be thinking, "I'm a DB2 for z/OS person. What's in the z13 for me?" In this blog entry, I'll give you my take on that question.
My favorite z13 feature is the larger and less expensive memory resource available on the servers. From a technical perspective, I like what I call Big Memory because nothing boosts application performance and CPU efficiency in a DB2 for z/OS environment like expansive real storage (as I pointed out in an entry that I posted to this blog a couple of months ago). A single z13 server can be configured with as much as 10 TB of memory (up from a maximum of 3 TB on a zEC12, the previous top-of-the-line mainframe), and a single z/OS LPAR on a z13 can use up to 4 TB of memory (with z/OS V2.2, or z/OS V2.1 with some PTFs) -- that's up from 1 TB for a z/OS LPAR on a zEC12. How much memory should a z/OS LPAR have? I will tell you that for an LPAR in which a production DB2 for z/OS subsystem is running, I like to see at least 20-40 GB of real storage per engine -- and I mean total engines, zIIP as well as general-purpose (so, for example, if a z/OS LPAR with a production DB2 for z/OS subsystem has eight engines -- four general-purpose and four of the zIIP variety -- then my recommendation would be to configure that LPAR with at least 160-320 GB of memory).
How would I want to exploit that memory for DB2 performance purposes? Well, for starters I'd like to have a big buffer pool configuration -- to the tune of 30-40% of the LPAR's memory resource (so, for example, if I had a z/OS LPAR with 300 GB of memory then I'd want the aggregate size of all the buffer pools allocated for a production DB2 subsystem in that LPAR to be 90-120 GB). I'd also want to have PGFIX(YES) specified for most of these buffer pools -- certainly for the more active pools. I'd also give consideration to specifying PGSTEAL(NONE) for one or more pools, and using those to cache some really performance-critical table spaces and or indexes in memory in their entirety. [Note: my 30-40% of memory guideline for the size of a production DB2 subsystem's buffer pool configuration assumes one production DB2 subsystem in the LPAR. If there were several production DB2 subsystems in a z/OS LPAR, you would not want each of them to have a buffer pool configuration sized at 30-40% of the LPAR's real storage resource. If there were multiple production DB2 subsystems in an LPAR, I would generally try to keep the combined size of all of the subsystems' buffer pool configurations at not much more than 50% of the LPAR's memory.]
I would not stop my exploitation of a z13 Big Memory environment with a large buffer pool configuration. In addition to that, I'd look at enlarging a production DB2 subsystem's dynamic statement cache, to drive the cache "hit ratio," ideally, to 95% or more (avoided "full" prepares of dynamic SQL statements can reduce CPU consumption significantly). I'd also go for a larger in-memory sort work area (specifically, I'd consider a value of SRTPOOL in ZPARM of 40-60 MB or more, and a MAXSORT_IN_MEMORY value of 20-30 MB or more). Additionally, I'd make greater use of the RELEASE(DEALLOCATE) package bind option, together with persistent threads (e.g., high-performance DBATs and CICS-DB2 protected entry threads), to improve the CPU efficiency of frequently executed programs. With the kind of LPAR memory I'm talking about, I think that I could do all of these real-storage-leveraging things and still have a demand paging rate of zero (though I wouldn't get too concerned if I had a small but non-zero demand paging rate of maybe 1 or 2 per second).
Now, I mentioned that I like the z13 memory picture from both an expanse and an expense point of view. I just covered the expanse angle. Now a look through the expense lens. The cost of memory on the z13 starts out substantially lower versus the cost of memory on a zEC12 or z196 server (the latter being the mainframe generation that preceded the zEC12). Depending on how much real storage you order for a z13, beyond what you have on a zEC12 or z196 from which you are upgrading to a z13, the memory cost can go lower still -- a LOT lower. Talk to an IBM z Systems sales representative for details about the memory deals available when upgrading your mainframe to a z13. They are impressive, if I do say so myself. Time to load up on gigabytes (or terabytes).
While I am definitely enamored with Big Memory (as DB2 for z/OS people ought to be), that's not all that I like about the z13. I'm also a fan of two enhancements that, in different but complementary ways, enable z13 processors to work smarter for enhanced application efficiency and throughput. One of these processing enhancements is called SMT -- short for simultaneous multi-threading. SMT allows multiple software threads to run on the same processor core at the same time. On a z13, two threads can execute concurrently on one SMT-supporting processor, and that is why the feature is sometimes referred to as SMT2. With SMT in effect, each of the two threads using one core will run more slowly than would be the case for a single-thread core, but throughput is boosted. My colleague Jim Elliott likes to use the example of a two-lane road with a 45 miles-per-hour speed limit versus a single-lane road with a speed limit of 60 miles per hour (and don't read anything into these speed limit numbers -- the only point is that one is lower than the other). Cars go faster on the one-lane road, but considerably more cars per unit of time travel a given distance using the two-lane road. In other words, the two-lane road with the lower speed limit has a greater carrying capacity than the one-lane road with the higher speed limit. Similarly, an SMT-exploiting processor should provide greater transactional throughput than it would if it were running in the traditional single-thread mode.
SMT is not applicable to all z13 engines. Specifically, the z13 processors that support SMT are zIIP engines and IFLs (the latter being an acronym for Integrated Facility for Linux -- processors dedicated to Linux workloads on z Systems servers, running either in native, "on the metal" LPARs or as virtual Linux systems in z/VM LPARs). That these engines support SMT is good from a DB2 for z/OS standpoint. zIIP engine utilization is often driven to a large extent by DB2 for z/OS DDF workloads (i.e., client-server applications that access DB2 data over network connections). DB2 for z/OS also utilizes zIIP engines for significant portions of utility processing, and (starting with DB2 10) for prefetch read and database write operations. Another reason z13 zIIP support for SMT is good for DB2: Java programs running in z/OS systems use zIIP engines, and z/OS is an increasingly popular environment for Java applications, and those applications very often involve access to data managed by DB2. z13 IFL support for SMT is great performance news for applications that run under Linux on z Systems, and many such applications interact with DB2 in an adjacent z/OS LPAR.
The other "work smarter" z13 processing enhancement that I like a lot is SIMD, or Single Instruction Multiple Data. With SIMD, if the same operation needs to be performed on several data elements, the operation can be performed once for all of the data elements at the same time, versus being performed for first data element, then performed again for the second element, then again for the third, etc. Again, fellow IBMer Jim Elliott provided a nice analogy: suppose you need to get three packages of the same type from point A to point B. It would be more efficient to do that by sending all three packages in one truck, as opposed to sending three trucks carrying one package apiece. Here's the DB2 for z/OS angle: SIMD should boost the performance of two kinds of application in particular: those that are data-intensive (referring to the volume of data operated upon) and those that are compute-intensive (referring to operations performed on data elements). Put those two application characteristics together, and what do you get? Analytics (among other things). In recent years, there has been a marked increase in the use of z Systems servers for analytics applications, driven in part by the steady convergence of transactional and analytics processing. z13 SIMD technology will add fuel to this trend, and it's a trend that most definitely benefits DB2 for z/OS. Software has to be modified to take advantage of SIMD, but that work is already underway. Look for SIMD exploitation by IBM analytics tools in the near future. And, in a z/OS V2.2 system (and V2.1 with some PTFs), SIMD will be exploited by XML System Services (used for things such as schema validation when XML data is stored in DB2 for z/OS tables), by our Java SDKs (good for WebSphere Application Server), and by our COBOL and PL/I compilers.
And that, folks, is the short and sweet summary of what I most like about the z13: Big Memory (like jet fuel for DB2, and priced to sell), SMT (like multi-lane highways), and SIMD (like shipping multiple packages in one truck). There's plenty more z13 technology that is cool in its own right, but as a DB2 for z/OS specialist I'm keeping my "big three" features front-of-mind.