Submitted by IBM has unveiled its latest mainframe, the System z10. The product is designed to improve performance and reduce power requirements and cooling costs. The new machine, based on quad-core technology, is equivalent, in terms of performance to nearly 1500 x86 servers, according to IBM. The z10 is designed to be up to 50% faster and offers up to 100% performance improvement for CPU-intensive jobs compared with its predecessor, the z9, with up to 70% more capacity, IBM said. IBM said the new mainframe would also consume 85% less energy and have a footprint that is up to 85% smaller.
I read somewhere that the 54-way z9, the top end setup, runs around $40 million. Anyone know if this is true? I have no idea if this would include disk or tape or if the writer was talking about just the machine and (?) the IBM software charges.
That would be about spot on.
The company I work for sells IBM mainframes to banks here in Luxembourg and I can say, without a doubt, that they incredibly expensive to buy as well as to operate.
so why do customers buy em? Because they can’t be beaten in performance and considering the amount of servers they can replace, they actually help reduce the total cost of IT operations.
against a Google-style megacluster of commodity servers? Power consumption is probably less, but I somehow get the feeling that the upfront costs are WAY higher for the mainframe. Still the idea of replacing a warehouse full of x86 servers with a single refrigorator-sized unit must be attractive to SOMEONE given that mainframe sales are at their highest levels since the 80s.
And given that the throughly entrenched businesses whose core operations all run on these expensive mainframes they’ve had for years and years and years would have to spend millions more than the cost of an upgrade to re-engineer all the software running on these systems…..
Nothing quite like the sweet aroma of vendor lock-in.
Lol! How true. Anybody who thinks MS has the copyright on lock-ins should know that they learned from the best, IBM.
There are some really cool things you can do with Mainframes though, like setup virtual servers and run Linux on em. I’d so love to see LAMP on a mainframe, totally kick ass! 😉
I -have- seen this!
At a former employer of mine, one of the linux-nuts (Ed, can you hear me?) talked one of the mainframe guys into creating a small LPAR in the test region. He had a fully functioning LAMP stack running.
While doing development on a centralized J2EE domain model, we used the DB2 install in a z/OS LPAR prior to moving to dedicated x86 servers running DB2 on Linux for production. Don’t ask. 🙂
In short, I’ve had more contact with big-iron than I’d care to admit to.
Edited 2008-02-28 20:45 UTC
Wow, I’m green with envy, you lucky dude!
I would so love to do some z series LAMP benchmarks. I’m pretty sure the sheer performance would be a webmaster’s wet dream.
After I’v finished my p series certs, I’m seriously considerings doing some z/os courses. Mainframes aren’t showing any signs of disappearing, in fact they seem to be getting more popular as time goes on. Surely it can’t harm having something like that on your CV 😉
yep, looking at some info over at wikipedia:
http://en.wikipedia.org/wiki/Z/Architecture
one can read that the Z series is backwards compatible with system/360, a series of mainframes that started out in 1964!
now thats some old code!
Remember when you replace a ‘WareHouse’ of servers with a fridge sized unit, don’t forget the cost/rent of the warehouse, the heating/cooling bill for a warehouse, the building maintenance cost of a warehouse, …. etc.
The freedom of being able to locate that ‘fridge’ anywhere that is safe can make for huge savings.
However, make sure it is safe. If anyone steals you entire IT department you are #$%&#&$.
Can this thing run FreeBSD 7?
it would not surprise me if it indeed could do so in some kind of virtualized/emulated system…
No.
Never say never … you could probably run a Linux distro in an LPAR, run KVM in the Linux instance, and then install FreeBSD into there.
So it’d be something like FreeBSD on KVM on Linux on LPAR on whatever OS the zSeries runs.
I know you’re joking, but that wouldn’t work. KVM allows you to run an x86 guest on an x86 host, and maybe someday it will run a zArch guest on a zArch host, but it won’t run an x86 guest on a zArch host. I guess QEMU might work, but that’s pretty pointless.
Oh yeah. Forgot about the i386/amd6-only-ness of KVM.
Well, there’s always bochs, qemu, virtualbox, etc. While it wouldn’t be ideal, it should (in theory) be possible.
Pointless, maybe, but possible.
This might be a somewhat ignorant question but: what exactly differentiates a modern mainframe from a more conventional (i.e. PC-style) server? I’ve never really heard a clear explanation of that and I’m curious.
>This might be a somewhat ignorant question but:
>what exactly differentiates a modern mainframe from a
>more conventional (i.e. PC-style) server? I’ve never really
>heard a clear explanation of that and I’m curious.
In short: bandwidth, reliability and precision.
The bandwidth out of these things is mind blowing, 150GB / second – and that’s only on one sort of link.
They have all sorts of reliability things for correcting errors pretty much everywhere, if a cosmic ray zaps a calculation it detects it and does it again.
Binary cannot represent decimal fractions exactly so these sorts of machines often do calculations in decimal, these machines now have decimal floating point.
You can also plug in all sorts of weird co-processors (including Cell), 1.5 Terabytes of RAM, god only knows how much disc space and all sorts of other weird expensive stuff.
They’re also CISC through and through, and its almost entirely done in hardware.
You pay through the teeth, and the nose, simultaneously, for them but the specs are pretty incredible.
Ironically though, in terms of raw computing power they’re probably not that powerful.
Edited 2008-02-28 22:12 UTC
Thanks for the clarification
Cell co-processors do not exist for the mainframe. That gaming company from Brazil was connecting cell blades to their mainframe with Ethernet. Their hope is to have a native cell attachment in the future.
Mainframe performance is a myth, it is something a lot of people talk about but none can make any substantiated claims simply because IBM will sue your ass as soon as you try to make any performance comparisons public. It is a proprietary platform through and through which has many more negatives than positives. Mainframes are the most costly computing platform known to man and I can’t say that it is justifiably so. You can get much better performance and pretty much the same reliability and availability from high end Unix platforms and will end up paying a whole lot less. This is especially true with z10 sharing a lot of silicon with IBM pSeries. You will get equal or possibly better performance on pSeries as opposed to mainframes, but it will be a lot less cheaper to run. You ask, so why would people still keep on using mainframes? The answer is inertia, the same reason people stick with Windows when Linux is clearly superior. Mainframe is way past its prime on technical merits for its asking price, but simply due to the nature of the applications running on top of mainframes and the simply phenomenally proprietary nature of the platform there is tremendous amount of inertia behind them. I can understand people buying mainframe simply because they are “stuck” with them, but otherwise if you’re considering becoming a mainframe customer take this piece of advice – take a gun and shoot yourself in the foot, it will hurt a lot less than being screwed by IBM every step of the way.
Here is a very nice little review on the mainframes in general:
http://blogs.zdnet.com/Murphy/?p=594
That ZDNet post has one thing wrong: System z doesn’t use Power5/6 processors and the z9 internals were quite different from Power5 systems. z10 and Power6 have converged more, although not as much as that post would lead you to believe.
http://realworldtech.com/forums/index.cfm?action=detail&id=83745&th…
Indeed the two basically share some common “DNA” such as caches etc but are quite radically different cores. Have a look at the ibm presentation on the matter:
http://www2.hursley.ibm.com/decimal/IBM-z6-mainframe-microprocessor…
Z6 is the old name for what is now the Z10. Here is a great article about the Power6 to compare with: http://realworldtech.com/page.cfm?ArticleID=RWT101606194731
Good articles. Thanks!
What really cracks me up is linux people enthusiastically boosting one of the world’s most proprietary platforms. I suspect that the perceived glamor outweighs the lack of freedom.
Now, I think that choice of platform is the perogative of any organization, whether the reasons are smart or stupid, so I am not going to criticize mainframe shops. I am a little suspicious of the wisdom of introducing the mainframe to an open systems shop, but hey, IBM salesmen need BMWs too.
However, the performance hype in the announcement is a little much for me, when swallowed hook, line, and sinker. Even if they are talking about 1500 single cpu x86 boxes, that would make each zSeries cpu 23 times faster than, say, a P4. Folks, if IBM has accomplished this, expect some nobel prizes in Armonk.
Truth is, they are talking about 1500 heavily underutilized x86 servers. That’s fine, but it would be more reasonable to say that the new box can replace “Somewhere between 50 and 1500 x86 servers, depending on if you like to use your CPUs or just spend most of your time waiting on users and disks”…sigh, I guess I’ll never get a marketing job…
Mainframes are not purchased for raw numerical performance – they are purchased for their rocksolid reliability where cost of failure and loss of service are considerable. The TCO factors in the cost of such failures (financial and reputational) along with the more usual savings in areas such as data center infrastructure and utility bills.
With a mainframe, you’re paying for an incredible piece of engineering with no compromise on component quality and incredible levels of reliability out-of-the-box. We’re talking run-time self-diagnostic fault isolation, fault-tolerant non-disruptive hot swappable components, execution integrity, in-flight processor migration from failing units, self-diagnosis and predictive deconfiguration of troublesome components, and the ability to remove *all* single points of failure — short of someone nuking the site from orbit (along with the backup data centers).
Taking one specific example, processing units have internal mirror functionality, effectively executing instructions twice and comparing the result.
Its just unfortunate that, in our throw-away “made-in-China” society, people are so eager to disparage such levels of engineering and the intellectual R&D effort that underlies it.
With IBM though, one thing that annoys me considerably is the kneecaping of CPU speed (IBM calls it Capacity Upgrade on Demand) using microcode to insert null cycles into the instruction stream until you pay for a higher workload license. A “Customer Initiated Upgrade” can actually be done online, and this opens up the possibility for “remotely initiated downgrades”. In fact, the system can automatically downgrade itself anyway as a result of certain expiration criteria.
> With a mainframe, you’re paying for an incredible piece of engineering with no compromise on component quality and incredible levels of reliability out-of-the-box. We’re talking run-time self-diagnostic fault isolation, fault-tolerant non-disruptive hot swappable components, execution integrity, in-flight processor migration from failing units, self-diagnosis and predictive deconfiguration of troublesome components, and the ability to remove *all* single points of failure — short of someone nuking the site from orbit (along with the backup data centers).
> Taking one specific example, processing units have internal mirror functionality, effectively executing instructions twice and comparing the result
As I have said before the features you pointed are not unique to mainframes, pretty much all of those reliability, serviceability, availability features are available at a fraction of a price in high end Unix system. Just look at the Sun SPARC M9000 or even the old Sun Fire F25k, the former can do everything you outlined above including memory mirroring and instruction retry. And you get these features on a platform that is running an open-source operating system (Solaris). Mainframe is way past its prime, the only reason it’s still alive is because of the inertia and proprietary lockin.