Honey, I shrunk the mainframe: Fujitsu freshens up GS21 kit Mainframes never die, though their architecture has been eclipsed for years. In Fujitsu's case, the GS21 2400 and 2600 mainframe models have become smaller and, we're told, faster and more secure with the arrival of the 3400 and 3600. The GS (Global Server) mainframes are sold mostly in Japan and have roots in Amdahl …
Interesting. Say more. I am a confessed greybeard mainframe bigot, and I understand that the world has passed me by... so looking for education on that 'architecture eclipsed' bit. If you mean market and/or mind share; obviously... but this phrasing implies technical obsolescence. So are you referring to the insane amount of redundancy/resilience built into mainframes? I work for a large company with tens of thousands of x86 and proprietary (AIX, Solaris) servers of every stripe imaginable; converged and plain, virtualized and not. Every morning on our operations call there are alway a handful that have had some failure of some ilk. Sometimes they fall over to their secondary, sometimes not. In my 8 years here, we have had exactly ONE hardware failure in our 30k+ MIPS mainframe collection, and the internal redudancy of the box allowed the workload to continue un-interrupted for a week while we analyzed the failure and scheduled a fix. Turned out to be a microcode issue. So is THAT the eclipsed architecture, or do I need to go into I/O capabilities?
or do I need to go into I/O capabilities
That's what PC people fail to understand, mainframes are not about heavy processing, that's better done on custom GPU jobs, but astonishing amounts of I/O which mainframes do far better than any system based on desktop processors*. And as you point out they have inherent resilience reducing downtime to a tiny fraction of other architectures.
There's room in the computing spectrum for many sorts of hardware, mainframes fill a niche which while smaller than it used to be still exists and will continue for a good while yet.
* currently and for the near future. Eventually big iron will be eclipsed, just not today.
The statement that big iron eventually will be eclipsed, presumably in I/O capacity, is rather badly in need of justification. What are the reasons to think that mainframe designers and vendors like IBM or Fujitsu will be unable to match the assemblers of giant stacks of Intel or ARM commodity processors?
Justification:
OK, I/O capacity will be hard to match in other architectures, but costs per I/O channel will converge so in some circumstances big iron will be hard to justify on economic grounds.
I doubt if big iron will ever go away completely but the number of installations will go down. Not replaced by Intel/AMD stacks but some intermediate design using the best parts of big and little systems, a design not as yet on the market but I'm sure the big vendors are working on it, not because they want to kill off big iron but they know if they don't provide suitable kit somebody else will.
So you had ONE hardware failure on the Mainframe, but many failures on the other servers? Well, if you had 10.000s of Mainframes you would have also seen dozens of failures every day. Sure, Mainframes are more reliable than all except high end Unix and OpenVMS systems, but you should not compare a few Mainframes to 10.000s of servers.
Regarding the cpu power of Mainframes - they are far more slower than x86. IBM tries to trick everyone to believe the z cpus are the worlds fastest, when in fact they are much much slower than x86 cpus. For instance, IBM claimed that one z10 Mainframe could replace 1.500 x86 servers. I dig into this and it turned out that all x86 servers idled and the Mainframe was 100% loaded. Well, in that case my laptop could replace 10 IBM Mainframes - provided they all idled. As one x86 cpu is much faster than Mainframe cpus, there is no chance that the Mainframe would be able to replace a few x86 servers if they started to do heavy computations. Mainframe cpus are slow.
However, Mainframes have superior I/O capability. That is the effect of lot of I/O co-processors that the Mainframe have. But OTOH, if you also gave the x86 servers an equal amount of I/O co-processors the x86 server would smoke the Mainframe in terms of cpu and I/O performance.
you should not compare a few Mainframes to 10.000s of servers.
If the throughput of a few mainframes is similar to rooms of x86 servers then it's a valid comparison. it's not just the failure rate but the failure mode that's important, when x86 devices fail they tend to stop working, when mainframes "fail" more often than not they send the admin a message and reroute around the afflicted part and continue working, a bit slower but without interruption.
if you also gave the x86 servers an equal amount of I/O co-processors the x86 server would smoke the Mainframe in terms of cpu and I/O performance.
Er, no. The x86 architecture is not designed to carry the I/O loading, if what you say was true then people would be doing it. While you can set up a lot of I/O channels on x86 you cannot use them all simultaneously at full bandwidth.
Regarding the cpu power of Mainframes
Yes you are completely missing the point of big iron, they are never used for processor heavy tasks, that is not what they are designed for. Simply they are for moving huge amounts of data around very quickly and really that's it and that's something x86 does not do well, yes one PC can move a fair bit of data around but it does not scale well and for massively parallel work, forget it.
It might also be pointed out that running your mainframes at 100% (or as near to) capacity 24/7 is exactly what customers want to do, otherwise your are just taking up space and generating heat that you have to pay to remove again.
And no, if you had 10,000 mainframes, you would not also see dozens of failures every day. They are not built like your bloody laptop. Now go away and stop talking shite.
There is a lot more to "mainframe" than executing the same instruction set and running the same operating system.
About 1993, a subsidiary of Gould challenged my (DoD) agency to match their cluster, based on Motorola four-core processor, bundles of memory, and a 70 megabit/second (!) backplane running Unix against our (at the time) IBM ES9000 running MVS/ESA. After some months and a good deal of preparation that included running without the "overhead" of ACF-2 because the challenger had nothing comparable, the contest was engaged.
I do not remember the entire result, but one chart from the report stood out. It showed response time vs. transaction rate for a simulated interactive application. On the challenger's system, the response was well under a second up to a point that was fairly high, after which it increased at an increasing rate until at the end of the test it was four or five seconds and rapidly increasing. The same test run against the ES9000 showed a response time of about 1.1 seconds. At all offered transaction rates. We kept the ES9000, which also offered reliability, availability, and service that the challenging system did not.
Mainframes are stunningly profitable. Code written in the 1970s is still wurdling around in the CPU and obviously those development costs were recovered long ago. The customer base is generally very happy to have something that works and the effort of migrating to something else is almost unthinkable anyway.
The only problem these days for mainframe vendors is that staff who know the hardware and OS inside out are near retirement age, and entry level employees would rather do something else.
Self identified balding greybeard here...
The IBM Mainframe Architecture allows 256 I/O channels. How many HBAs can you cram onto on x86 box? Or maybe better considered as, how many parallel I/O operations does your biggest favorite box support?
I work for a DASD vendor [spinning rust for the open systems types], and see MF boxen in the largest transaction processing environments, airline reservation systems, credit card authorization systems, supermarket cashiering apps, etc. I've watched several large customers try to replace the mainframe, but fail because they can't replace the I/O parallelism capability that the MF provides.
Until Intel, PCI or HBA vendor figures out how to match that type of parallelism, the MF will live on.
Random thought. I grew up when what we do was called Data Processing. We apparently are Information Technology today (or pick whatever label floats your boat). Arguably, if what you're trying to do is process data, I/O capabilities become pretty important.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
