Aug 6 ’06
Sweet New Mainframes
IBM has always been a little unsure of how to deal with the low-end mainframe customer. From the days of machines such as the 4321/4331/4341 (.18 to .81 MIPS!), through the 9370 series (the up-to-3-MIPS “VAX killer” that didn’t), to the Personal/390 and Multiprise systems, low-end systems allowed IBM to “plant acorns” that would grow into mighty oaks—towering 3084s, 3090s, and 9672s—as a shop’s processing needs grew.
But, to some extent, these systems always seemed like afterthoughts: They would run mainframe applications, but some features available on “real” mainframes— multiprocessors, vector support, PR/SM Logical Partition (LPAR), et al.—weren’t available. The Multiprise 3000 offered the closest thing to the “real S/390 experience,” including LPARs and even Integrated Facility for Linux (IFL) engines, but was limited to two processors and was structurally different from its big brothers.
In 2002, the z800 (and, in 2004, the z890) were introduced as the replacement low-end machines. These were “real” mainframes in most senses, although the z800 was built by Hitachi (with some IBM parts) and was physically quite different from the z900. (The z890 was much closer to a z990, using the same processor “books,” but was air-cooled and limited to one book per system.)
With the advent of the z9-109 last July, the obvious next question was, “Where’s the z8?” In late April, IBM answered this question, and simultaneously replaced the unwieldy “z9- 109” name. The new, low-end systems are now “z9 Business Class” (“z9 BC,” machine type 2096), and the high-end systems are “z9 Enterprise Class” (“z9 EC,” still machine type 2094). The BC machines are “real” z9s: Like the z890, they use the same books as their larger cousins, and are air-cooled and limited to one book per system, but mechanically they’re otherwise almost identical.
Long-time IBM customers will find the new naming confusing, to say the least. IBM spent the better part of two decades pushing the term “enterprise” as a synonym for “business”; in addition, Basic Control mode (BCMODE) and Extended Control mode (ECMODE) were hardware-level bit settings on the Sytem/370 that controlled whether the machine operated as a System/360 (BCMODE) or System/370 (ECMODE), so “BC” and “EC” were already “taken” in graying mainframers’ minds.
Nevertheless, the z9 BC is a great step forward for mainframe customers. Code-named “Pollux” in development— from Greek “Polydeuces,” which means “very sweet”—these are sweet boxes indeed. The z9 BC ranges from 26 MIPS (for a detuned or “kneecapped” single processor) to something on the order of 3,000 MIPS for a full, seven-IFL machine. For those of us who “grew up” with systems below 1 MIPS, it’s nothing short of amazing to see that a full-speed z9 engine is north of 500 MIPS, making it faster than a five-way 9672 G5 machine—some pretty serious iron not too long ago!
While it’s easy to shrug and say, “Yeah, and my Pentium 4 is 500 times as fast as my 8086 was,” the difference is that IBM has evolved the rest of the mainframe in a much more balanced way than PCs. Memory size, number of CPUs, and I/O bandwidth must all advance for raw CPU speed to be fully beneficial. The mainframe’s ability to support multiple tasks doing real work while running productively at 100 percent CPU utilization is something current Intel architectures can only dream of.
As a “real” mainframe, the z9 BC supports the same processors as the z9 EC:
- CPs—“regular” CPUs, capable of running z/OS, z/VSE, z/TPF, z/VM and Linux
- IFLs—CPUs that can run only z/VM and Linux
- Internal Coupling Facilities (ICFs) for z/OS Sysplex
- System z Application Assist Processors (zAAPs) (Java engines for z/OS)
- The new System z9 Integrated Information Processors (zIIPs).
Conceptually similar to the old Teradata DBC 1012 Data Base Computer, the zIIP provides database offload capabilities for DB2 work, but using standard databases residing on z9 disk subsystems. Essentially, a zIIP is a regular CP that doesn’t appear when the number of processors is queried; DB2 will tell the z/OS Workload Manager (WLM) to run various work on a zIIP, if available. We can expect to see other uses of zIIPs in the future beyond just DB2. (For a more detailed look at zIIPs, see “zIIPing Along,” by William Favero in this issue of z/Journal.)
The zIIP is a fully transparent solution and solves the same problems as zAAPs, reducing the number of “real” CPUs in a system and the cost of product licenses. All the specialty engines also give IBM added pricing flexibility: priced lower than CPs, they enable growth in specific areas (i.e., Java, DB2, Linux) without risking z/OS revenue. With On/Off Capacity on Demand (OOCoD) for all these processors, a single machine can expand and contract as business needs dictate.
Processor configurations for the BC are somewhat confusing. You can have up to four CPs and up to seven IFLs, up to the limit of seven total processors available on the machine. The number of zAAPs and zIIPs is limited by the number of CPs: You can’t have more zAAPs than CPs, nor more zIIPs than CPs. With two CPs, you could have up to two zAAPs and up to two zIIPs; with three CPs, you could use all four of the other engines as some combination of zAAPs and zIIPs, as long as you had no more than three of either; and so forth. Other interesting aspects of the April 27 announcement include:
- FICON Express4, which supports transfer speeds of up to 4Gb per second at distances of up to 10km, continuing the mainframe’s “further, faster I/O” growth
- z9 HiperSockets now support IP V6
- N_Port ID Virtualization (NPIV) lets a single FCP port register multiple Worldwide Port Names (WWPNs) with a fabric nameserver, thus allowing a single FCP port to appear as multiple WWPNs in the FCP fabric
- z/VSE 4.1, which finally moves VSE into the 64-bit arena.
The speed at which “obsolete” IBM mainframe technology is evolving is dazzling. New machines, new processor types, new connectivities … and even some analysts who made dire predictions about the death of the technology are grudgingly admitting they were not only wrong, but wildly wrong.
Besides basic technological enhancements, new uses for the architecture— Linux, virtualization in general, and Service-Oriented Architecture (SOA)— are all areas in which mainframes excel. SOA is really just client/server and middleware “done right,” with the important data residing centrally and the presentation handled locally, thus taking advantage of the respective strengths of mainframe and Intel hardware. And it’s a lot harder to lose a mainframe containing your customer data than a laptop!