More Insight into IBM's New z114
Rob Enderle's reaction to IBM's introduction of the new z114 mainframe is quite interesting and thought-provoking. In short, he notices that major IT vendors have flipped roles, with Oracle (and Microsoft, I would argue) now emulating IBM's worst behaviors from the 1980s, undermining trust among formerly loyal customers and partners. Enderle also observes something I reported in my reaction to the z114's introduction, that IBM is passing along its cost savings (and perhaps more than that) to its customers — clear evidence that IBM is focused like a laser on growing the mainframe ecosystem. Said another way, Sun customers are to be milked, and IBM customers (current and prospective) are to be delighted.
Another way IBM is delighting current and prospective mainframe customers is on the software side. As most people in IT purchasing know, software licensing and maintenance fees are increasing quickly, in general. I point to these key drivers:
- Commercial software is a close substitute for labor. (You can either buy software or build it.) Given long-run global trends in labor and in business sophistication, we would expect commercial software to take a progressively bigger role in IT projects with correspondingly bigger shares of the project budgets.
- There has been quite a bit of consolidation in the software industry generally. Less competition tends to result in higher prices. Interestingly, there has been increasing mainframe software diversity in recent years.
- Virtualization tends to drive down the number of CPU cores required to perform the same work. Software vendors have been racing to claw back their licensing and maintenance fees as customers have virtualized their software products, most of which are licensed per core.
- In some industries there's unexpected growth that's driving additional infrastructure expense (including software). For example, I'm hearing from a lot of banks that smartphones (such as Apple's iPhone) are driving up transaction volumes and infrastructure requirements.
- As software becomes more sophisticated, new entrants face tougher challenges getting established. If there's less competitive threat from potential new entrants, then prices will tend to be higher.
That said, IBM is consistently driving down mainframe software unit costs, something that's quite rare or even unique in the industry. The z114 continues that evolution, with even more attractive software licensing terms. IBM dropped prices for z/OS and other mainframe-unique software products. You'll probably see a reduction in z/OS-related software charges ranging from 5% to 18% just by moving from a z10 BC to a z114. Interestingly, IBM's most aggressive z/OS pricing is for single z114 customers. Coincidentally my favorite mainframe configuration is the "single machine virtual cluster" Sysplex approach, and IBM's new pricing is especially friendly to that mainframe-unique, parsimonious way of delivering exceptional reliability and availability.
At the same time, IBM dropped its prices for its Linux-based software products. Pricing per core fell from 120 "value units" to 100, a decrease of almost 17% compared to the z10 BC. Core performance is up substantially, so that means IBM reduced unit prices roughly 29% (by my quick calculation) compared to the z10 BC.
There's no question this IBM behavior is customer-friendly, and it goes a long way toward explaining why IBM is enjoying such phenomenal mainframe growth. I hope more IT shops understand this stuff, and soon.
|by Timothy Sipples||July 22, 2011 in Blogs, Economics |
TrackBack URL for this entry:
Listed below are links to weblogs that reference More Insight into IBM's New z114:
I have to say my impression has been that mainframes are terrible for Internet services.
For systems like those that keep track of bank account information, hotel reservations, flight bookings, etc., mainframes are a great fit.
But those systems typically are front-ended by web services, and it seems to me those web services don't seem very well suited for a mainframe. The relative compute power, throughput, and transaction rate of a mainframe per rack unit and power unit seems to be paltry compared to a rack of recent x86 blade systems. Did I miss something?
I've run (31-bit) Linux on an LPAR at one point at an IBM class in about 2006, and I found it exceptionally slow. I'd hate to imagine running some sort of application server on a mainframe. You'd need a lot of mainframes I would think.
Posted by: Tony Bourke | Aug 10, 2011 10:54:02 AM
That was a long time ago, Tony. If that class was similar to the class I attended about that time, you were using a pre-Linux mainframe from the late 1990s called a Multiprise 3000, which is roughly the size of an American home washing machine. (If you were running 31-bit Linux in 2006, that's a strong hint it was a Multiprise 3000.) That entire machine topped out at about 200 MIPS, but your class might well have been running at 60 MIPS (H30 model). It ran OS/390 and VSE/ESA "OK" but was never a good Linux system, even when it was new. (IBM positioned the MP3000 as Linux-capable but not Linux-performing. It was a Linux "sandbox.") The disk was slow, the I/O was clever but slow, the cache was reduced, the memory was tiny (and slow), the clock speed was dialed down -- IBM designed that machine strictly for compatibility and affordability, not for performance or even (really) throughput.
And it was a great machine -- don't get me wrong. It's one of my favorite mainframes in history. But you have to be aware what it was designed for. That background helps explain why there were some surplus MP3000 machines available for lugging around to classes.
IBM probably told you "don't judge performance on this," and rightly so. (Although a class full of students all sharing a tiny system like that is still pretty amazing.) But unfortunately most people forgot that caution.
For comparison, a single z196 machine supports over 52,000 MIPS, and each MIPS is much improved for Linux compared to that MP3000. The smallest Linux mainframe that IBM sells (the z114) doesn't have a published 10-way MIPS number but is probably 5,000+ MIPS, also with the big per-MIPS performance boosts since the MP3000. Interestingly the starting price for that z114 is much lower than the MP3000's starting price, even in nominal dollars.
That said, if you're looking for *cheap* computing, with everything that word implies -- calculations-per-dollar (acquisition price) -- check out X86 servers from Intel and/or AMD. IBM sells those, too. (Or perhaps even Intel Atom or ARM servers.) If you care about qualities of service, and/or you have massively virtualizable workloads, and you still want good-to-excellent performance, check out a modern mainframe.
Posted by: Timothy | Aug 10, 2011 8:55:04 PM
I'm sure there are a lot of companies that wouldn't mind spending extra money for reliability, but I'm looking at the MIPS comparison, it's actually orders of magnitude more expensive.
According to Intel, a *single* new Xeon E7 processor gives you about 96,000 MIPs. Even if Intel exaggerates by a factor of two, that's still one processor = 1 mainframe. I can cram 96 of those Xeons into a single rack, meaning one single rack = 96 mainframes.
I haven't done the math for power consumption, but I'm guessing the watt per MIPs would be much more favorable with x86.
Also, where does the mainframe advantage come from? The advantage of a Mainframe come into play? Is it the hardware?
With Linux, it wouldn't be the OS, since it's Linux. Linux doesn't have any stability problems on x86, so that's not any benefit. And the software that would run on Linux (various application servers) would primarily be limited by CPU, not the mainframes strong suite. I wonder how many mainframes it would take to replace a single rack of x86 blades.
Disk I/O is usually limited by IOPS. Spinning rust is the limiting factor these days, no matter what protocol you run it over. SSDs help by ratcheting up the IOPs to tens of thousands.
Posted by: Tony Bourke | Aug 10, 2011 10:09:14 PM
Also, I didn't meant that response to sound confrontational, I've just written off mainframes a long time ago, and I'm genuinely interested to figure out if mainframes are something I'm missing out on.
Posted by: Tony Bourke | Aug 10, 2011 10:27:52 PM
Well, you could try replacing one (modern) mainframe with one Xeon processor to prove that hypothesis. Do you think that would work?
I don't know if those MIPS numbers are accurate. But there are at least a couple problems:
1. If you tell an Intel CPU (or any CPU) to add the same two numbers (in registers) over and over again, you'll get a MIPS measurement. It won't be meaningful, though. An add instruction (for example) is only one type of instruction. Mainframe CPU instructions are the most complex. A single mainframe CPU instruction can wash your car. OK, I exaggerate, but only slightly. So you cannot compare MIPS across CPU types -- even X86 to NVIDIA GPUs.
2. Mainframe CPUs have a lot more support for staying "well fed." Things like gigantic caches, for example. Otherwise the CPUs wait for data and don't actually do any work. Virtualization is especially demanding -- something mainframe CPUs are tuned for over the course of decades. That's not the Intel/AMD heritage.
But I'm certainly not opposed to X86 servers. They're darn useful for many types of workloads.
Posted by: Timothy | Aug 11, 2011 12:29:23 AM
The comments to this entry are closed.