IBM Leaks Details on New Mainframe

Mark writes "Big Blue inadvertently revealed details about its new z10 Enterprise Class mainframe set to launch on Feb. 26, as well as details on z/OS v1.10, a new version of the mainframe OS due out in September. 'According to an internal IBM document obtained by SearchDataCenter.com, the z10 Enterprise Class will come in five different models and feature 64-way chips, compared with the 54-way z9 mainframes and earlier 32-way models. In a conference call last month, IBM CFO Mark Loughridge told investors that the z10 would have 50% more capacity, which indicates that it will probably tap out at around 27,000 million instructions per second (MIPS) at the top end, compared with about 18,000 MIPS on the previous z9 Enterprise Class.'"

Comments

Nah

[dwalsh]

I am not buying one till they get that OS up to 3.0 at least.

Imagine...

[palegray.net]

Just imagine a beowulf cluster of these. It would make my head explode.

Re:Imagine...

[mcpkaaos]

Just imagine a beowulf cluster of these. It would make my head explode.

Okay, but you asked for it.

Re:Imagine...

[quanticle]

Splitting one of these mainframes into multiple Linux VMs is actually one of the more common uses for these things.

This might be a dumb question...

[TubeSteak]

How come they talk about thousands of MIPS instead of just saying GIPS?

Re:This might be a dumb question...

[avalys]

MIPS stands for millions of instructions per second, not mega-instructions per second. We'd have to talk about billions of instructions per second, or BIPS, and that sounds lame.

Re:This might be a dumb question...

[QRDeNameland]

Actually "thousands of kilograms" would be "megagrams", but we generally call them "tonnes".

And from there it goes to kilotonnes and megatonnes, then I believe a thousand megatonnes is then commonly called a "shiteload" or, in the US, a "fuckload".

Re:This might be a dumb question...

[bipbop]

I resemble that remark!

Re:This might be a dumb question...

[mcpkaaos]

Allow me to offer a solution, inspired by our friends in the storage industry:

1024 MIPS = 1 GIPS
1000 MIPS = 1 GiPS

Re:This might be a dumb question...

[ozbird]

I believe a thousand megatonnes is then commonly called a "shiteload" or, in the US, a "fuckload".

A U.S. fuckload is a thousand megatons. A thousand megatonnes is a metric fuckload.

Re:54 way chips?

[cruff]

Nah, just a bunch of dual core chips. Take a look at the IBM Journal of Research and Development which has a lot of nice detail. Look at Vol. 48, No. 3/4 and Vol. 51, No. 1/2.

Re:Low-End Port to PowerPC?

[TheRaven64]

The z6 isn't very similar to the POWER6. In the most important aspect with respect to porting software, namely the instruction set, they are completely different. The z6 instruction set is an incremental improvement on an architecture that goes back all the way to the 1960s with System/360 - the longest running architecture to maintain backwards compatibility. The POWER6 architecture is an incremental improvement on an architecture which dates back to the mid '90s and was designed from scratch around a completely different set of ideas.

The things they share are not visible to the user as they are hidden behind the instruction decoder. You can see some evidence of the fact that IBM are trying to lower costs by sharing a lot of the design between the two lines though from certain new additions to the POWER instruction set, such as hardware support for Binary Coded Decimals (useful in high-throughput financial systems and present in the mainframe line since the 1401 and 700-series, which preceded System/360).

Re:Kinda slow, eh?

[palegray.net]

Considering a single modern quad core Pentium has about twice the processing power as this mainframe. You are kidding, right? These systems are massively parallel machines, and are frequently used these days to present dozens of operating system images running concurrently. They support nifty ideas like instant failover and clustering on one machine, with arbitrary SMP to scale up performance as necessary.

Re:Kinda slow, eh?

[TheRaven64]

Considering a single modern quad core Pentium has about twice the processing power as this mainframe Uh, what? Currently, the fastest CPU you can buy is the POWER6. This mainframe uses the the z6 CPU, which is effectively a POWER6 with a different instruction decoder and MMU, and it supports up to 64 of them. They are connected via an SMP hub chip which adds 24MB of shared level 3 cache.

They also support partitioning on the hardware level, so you can run z/OS or Linux virtual machines with almost no overhead (something you've been able to do since it was called System/370). You also have a huge amount more fault tolerance with a system like this (take a look at how many transistors on the CPUs are dedicated to error checking, and then start looking at the peripheral systems).

Re:Kinda slow, eh?

[dreamchaser]

You don't get out much do you? Mainframes are going strong in data centers that need high availability, fault tolerant, error correcting, massively parallel systems. There is also a LOT of old code that is still going strong on them. Their inherent ability to run multiple virtualized OSes is another strong suit.

Your math is also way off if you think 4 x86 cores outperform this. I'll leave you to do the proper calculations as your homework.

Re:Kinda slow, eh?

[Marnhinn]

I work currently in an organization that uses mainframes. They are the z9 series - and honestly are some of the most useful things.

We use them due to ability operate in something called a sysplex. A sysplex is when multiple mainframes share data (known as DASD) and work together. When a mainframe is in a sysplex, you can do all sorts of things to the machine without having to bring your application down. These range from whole operating system upgrades to hardware maintenance and the end user will never see the impact. A sysplex literally is designed to be a 24x7 operation.

You can buy other types of machines that will be more powerful, faster or do operation x better, but it is hard to find a set of machines that are as stable and reliable as a mainframe is (and process millions of transactions per second).

Also, in terms of virtualization - a single mainframe on z/Linux can host many virtual linux servers - enough that you can save a substantial amount on power costs (my org estimated 400k a year in savings in terms of power alone - if the linux servers that are hosted individually on one of our distributed networks went to virtual on a mainframe).

Re:Nah Dried off?

[davidsyes]

I'll wait til it's dried off. Can't have a server that someone took a LEAK on...

Oh, maybe they made early RELEASE of details... I wonder, in IT context, how a vendor can "leak" its own details...

Re:Kinda slow, eh?

[dedazo]

The day you can use commodity hardware to build a failover-capable sysplex running multiple instances of an OS that can run 30-year old COBOL applications that do millions of financial transactions per minute with absolute 24/7/365 uptime, you'll be a very rich man indeed.

In the mean time, IBM, Hitachi and a few others will be raking it in for you.

And we just upgraded last month

[jocknerd]

I guess the IBM rep forgot to mention that a new one was one the way.

Re:Except the processor...

[TheRaven64]

I've seen that before, but the POWER6 is not the convergence chip. They share a lot, but it's stuff that matters from IBM's perspective (lowers development costs), not from their customers' perspectives (allows them to run the same software). That said, the biggest difference between the POWER6 and the z6 is the instruction decoder (addressing modes are easy to switch - look how many the Core 2 supports). Instruction decoders tend to take a roughly constant number of transistors. One of the big wins for RISC chips was that they could have simpler decoders and dedicate a lot more transistors to execution units. Now, the decoder has gone from being 50%+ of a CISC chip to under 10%.

It wouldn't surprise me if the POWER7 either implements a superset of the POWER and System Z architectures, or has switchable decoders. Considering the fact that it's already possible to hot-plug CPUs on systems at this level, I can imagine a future IBM line where the hypervisor allows you to not only partition the system, but also decide which chips run in POWER and which in System Z mode dynamically, migrating virtual machines and restarting CPUs as required. That could be very attractive for customers wanting to consolidate mainframe, AIX, and Linux systems.

One of the design goals of the PowerPC instruction set (a superset of which is implemented by the POWER6) was to easily emulate x86. It would be really interesting if IBM would enhance this emulation support into the hypervisor, allowing customers to run legacy x86 Linux, Solaris or Windows Terminal Server virtual machines on their mainframes.

By the way, this mainframe is one of the big reasons why IBM are so keen on open source. If you run Linux and (portable) open source software then IBM can sell you a mainframe running Linux VMs when you start to outgrow your current infrastructure. The reason IBM owns so much of the (small, but incredibly lucrative) mainframe market was that in the '60s they pushed the predecessors to this system - System/360. They sold cheap minicomputers and high-end mainframes that ran exactly the same applications (and, with System/370, the mainframe could even run virtual minicomputers). They got people using the cheap minis and then presented them with a clear upgrade path. With open source, they can give people a really long upgrade path starting at commodity hardware and going as far up as they want.

Re:Kinda slow, eh?

[edibobb]

Z6 is nothing. I've used a Z80 before.