isham research
(This discussion has been overtaken by events. The current version will be found here)
IBM's pTerodactyl provides new low-end performance points.
At the bottom edge of the performance matrix, IBM has clearly responded as best it can to the emulation providers with n-way low power machines - but it has also provided machines that map very well onto the established operations of 3090 and earlier CMOS n-way users.
The smallest pTerodactyl of all - (Capacity Setting 110 - ca. 27 MIPS) - is configured differently to all the others. It supports fewer LPARS - 15 against 30 - and also fewer channels - 256 against 512 - compared with the other 27 systems. Whereas the other systems use eWLC, the smallest of them uses zELC. Overall, it is very similar in configuration and Ts&Cs to a small z800. For logistics reasons (sparing) it is highly unlikely that the 110 is any different from the other 27 systems at a hardware level - the difference could be achieved quite simply by a microcode change to disable a memory bus adapter.
The reason for this drastic degrade is probably not simply marketing - it's more likely to justify especially low pricing. In fact, it doesn't present a problem for potential users - the system is still if anything over-configured for its 27 MIPS. With the caveat, of course, that the "110" descriptor only refers to the z/OS workload - such a system might have two 366 MIPS IFLs and one 366 MIPS zAAP as well. Only one fortieth of the compute power of such a system is described by its Capacity Setting.
Although IBM heavily preannounced this system at the beginning of February 2004, it still does not have a viable solution at the very low end. This is a shame - although low end customers only provide a small fraction of IBM's mainframe revenue, albeit at good margins, they provide critical mass for smaller ISVs who significantly support the mainframe's viability. The z890's low end n-ways cover a range of 27 to 99 MIPS - the level at which emulation providers such as Cornerstone Systems and T3 Technologies and their business partners around the world have been submitting Special Bid applications.
There is a disconnect, however, between IBM's and Intel's processor strategies. IBM makes major announcements in each mainframe sector - medium and large - around every two years. In medium and large accounts, this is sufficient - zSeries has excellent horizontal scalability both within a system and within a Parallel Sysplex.
But Intel ships a new clock rate every few months. So if the low end z890s are positioned against emulation, they have very much taken on a moving target. It takes around 65MHz of Intel 32-bit performance to achieve one zSeries 31-bit MIPS. Emulation can easily provide 27 MIPS with only a 1.8GHz Intel processor - something available on IBM's own T30 ThinkPads since August 2002. That's eighteen months before the z890's announcement - and may tell us when IBM's design decisions were made. With IBM now shipping 3.0GHz processors, emulation is now approaching 55 MIPS per engine - although degraded (or "kneecapped") versions are also available, as with the z890.
And the Intel space moves fast. "How fast are they this week?" is a common question, and it is often hard to source a system for delivery with the clock rate that was bid to a customer only weeks before - Intel (specifically, the IBM xSeries) always seems to have moved on. Asking a broker for the slowest system he has is a novelty.
Just hitting the right performance level is only a part of the solution. Emulation also offers the opportunity of mixing and matching architectural level sets, permitting users to take their time migrating from older operating environments that won't run on IBM's new hardware. This also helps low-end server consolidation, where workloads may be gathered in from obsolete mainframes located in far corners of a large company. It is useful for ISVs, too, who often have to support their code on operating systems that will not execute on current hardware - Architectural Level Sets seem to be becoming more frequent.
Emulation also provides integrated DASD at a much lower cost - and higher performance - even than IBM's newest external storage devices. IBM has not offered either internal DASD or even internal controllers on small systems since the Multiprise - a break with a tradition stretching all the way back to the 360/25. It's possible that the cancelled "Flagstaff" system (Multiprise 4000?) would have had such a feature. The z890 has no internal DASD; traditionally a requirement for success at the low end and a major product weakness. It makes no sense to acquire a low end processor if the external storage devices cost three times as much.
The same applies to SNA - although there aren't many 3745s left, being able to emulate one on-board means an emulation user can dispense with both the 3745 and its NCP - a significant cost saving. Recently IBM has announced a Statement of Direction to support NCP functionality on zSeries under Linux - the first example, perhaps, of emulation leading IBM technology as Amdahl's MDF once did.
It's even possible to automate pseudo-tape handling - xSeries disk storage is now so cheap that many classic mainframe shops can keep their "tape" files on disk, and significant levels of automation can be achieved.
There is another side to the coin - a "genuine" mainframe can do many things that emulation cannot - e.g, support Coupling Facilities and hardware cryptography - features required even by some ISVs developing in this space. It also offers IBM's Capacity Upgrade on Demand and can attach much larger external peripheral farms. IBM will offer special rebates on z890s to qualifying ISVs.
Blind attempts to force the z890's 27-MIPS 110 on the market will result - as with the z800's 2066-0E1 - in massive damage to the low end installed base for very little return. No matter how much the price is shoehorned and the IBM software costs grandfathered, it's the wrong machine with the wrong technology for the market segment - telling IBM mainframe users to "jump" often makes them jump to Sun.
IBM has also failed to understand that - despite its best efforts - few ISVs in the low end space use MSUs for charging, and still fewer subscribe to IBM's "charging" vs. "performance" view of MSUs. Down here, software groups (abandoned by IBM on 1 January 2000) still determine many charges - and moving up even one group is absolutely impossible for many if not most small shops. None of the last few years' initiatives - PSLC, WLC, LPAR pricing, eNALC, "charging" MSUs have any traction in this space at all - and it's interesting to note that IBM's own LSPR still defines Software Groups for the z890.