IBM Launches Z13 Mainframe
www-03.ibm.com
www-03.ibm.com
- up to 141 configurable processors
- new 22nm 8-core processor chip@5GHZ
- 110GIPS
- Single Instruction Multiple Data (SIMD) support
- On chip cryptographic and compression coprocessors
- up to 10 TB of memory (configured as 'memory raid', RAIM)
PDF is here: http://public.dhe.ibm.com/common/ssi/ecm/zs/en/zsd03035usen/...
> Single Instruction Multiple Data (SIMD), a vector processing model providing instruction level parallelism, to speed workloads such as analytics and mathematical modeling. For example, COBOL 5.2 and PL/I 4.5 exploit SIMD and improved floating point enhancements to deliver improved performance over and above that provided by the faster processor.
Most people would consider COBOL and PL/I to be laughably out of date languages from the 60s and 70s. But here's a new machine with 10 TB of RAM, and the release notes are quick to point out new features for COBOL and PL/I. Go figure.
Seriously, we have code whose originals can be traced to systems long dead. Yet code is not stagnate, full SQL integration exists for most language, there are many ways to get to the web. That and you would be surprised at the number of companies who use these machines simply for one reason above all, reliability. Our last crash was a site failure and even then we rolled to the DR site instantly.
I didn't have to write lots of Cobol, but it's been very educational: both its very rigorous syntax (and number handling as you said!) and the fact that the code I was modifying had been written even before I was born.
"If it ain't broke, don't fix it"
Cobol in itself is not a bad language at all! Mainframes are obviously rock solid and for development purposes, open source implementations of MVS exist (and were already useful 15 years ago, such as Hercules[1])
Well they are. These things are built for a specific audience and it so happens that they are still willing to spend money on this, otherwise it wouldn't be made, especially not by IBM that is otherwise quick to disband business units that don't make heaps of money.
In my research I'm using an even older language: Fortran. Believe it or not, but in a field that is less narrow and specific than you would think[1], there's still no real replacement for (modern) Fortran.
[1] numeric applications that need to get close to the hardware's peak performance, written in a higher level language than C that is usable by non-CS-graduates.
Not to mention that most compilers have OpenMP suport built-in. Also, there's the tons of heavily-optimized Fortran math libraries that make it easy to take advantage of big machines and big clusters.
Those scientists that are still using Fortran after all these years know what they're doing. It's not that they're ignorant of new technoglogy. A well-informed and open-minded CS person would probably end up using Fortran for these use cases, too.
Yes. In my experience, you can make C as fast as Fortran, but it will be a lot more work (and you need more knowledge about the machine). Fortran's defaults are very strong when it comes to crunching numbers as quickly as possible. Multidimensional arrays with Matlab-like slicing and operations, implemented as performant as possible, is enough reason alone to stay with Fortran.
> Those scientists that are still using Fortran after all these years know what they're doing. It's not that they're ignorant of new technoglogy. A well-informed and open-minded CS person would probably end up using Fortran for these use cases, too.
Well, I'm sort of coming out of CS (computer engineering with focus on software) and I'm using it, so there ;)
So that leaves language level support. So far the only contender that I can see coming up is Julia - and it will take a lot of effort until it can reach performance levels on par with Fortran. This basically would require a big push by one of the big software companies - who all aren't making a lot of money in HPC software anymore. I could see Nvidia picking up the ball at some point, it would suit them well - but then we're getting into vendor lock-in again.
http://en.wikipedia.org/wiki/Row-major_order explains it nicely.
If you wanted to take a native Fortran matrix and turn it into a native C matrix, you'd either need to copy everything to re-arrange the in-memory layout or do funky things with the indexing. Neither option is great.
---
Bringing it back to a higher level, I'd agree that your point makes sense most of the time. In most cases, the cost of a memcpy(3) would be a rounding error. That's certainly the case for most web applications. In a tight loop in a math-heavy application, though, the standards are a bit higher.
In theory, sure. In practice a bunch of languages have tried and yet Fortran is still the king.
I'm sure optimized libraries could be written in any language, however the vast number of researchers have decided that Fortran is by far more accurate and faster (for these libraries) so they go with that.
Well, at least in anything that has really performant compilers. That's currently true for C, Fortran and Assembly, to some extent also C++. Between the first three, Fortran is by far the easiest to use for numerics - its syntax is in fact not much more difficult than Matlab. It's clunky in anything but numerical code (such as when you try to create a nice API or when you try to have generic methods), but at the end it still wins.
PL/I is not totally out of date! E.g., it has some really sweetheart scope of names rules. The On Conditions are generally nicer than Try-Catch, e.g., if in a code block A and do a Goto out of it to a statement label in a containing code block B (it's fair to call PL/I a block structured language), then the stack of that task (in some ways nicer than just a thread) is rolled (popped as in popping a stack, which is essentially what happens) back to the context of the last time block B was active.
Block B can have a function F that inherits names defined in the blocks that contain B, and block B can call a function G, pass F as an entry variable, and function G can call it's parameter for function F, function F can execute and access all the variables it has access to that function G may not. That can be nice. I used it once when scheduling the fleet for FedEx.
Can have tasks, like threads. Can allocate some storage as, say, controlled. Then when the task ends, that allocated storage is freed. That can be nice -- nice way to stop memory leaks when have to kill a task.
PL/I has structures which replace much of the utility of classes. And the structures, with arrays of structures of arrays of structures of arrays, etc., are much faster than instances of classes because all the addressing is just a fairly nice generalization of simple old array addressing without pointers.
Such a structure can be based which means that can allocate one and get back a pointer to it. So, that is a lot like allocating an instance of a class. B can be part of a structure A, and B can be defined to be like structure C -- so, can get some inheritance.
The flat file I/O is much sweeter than what is in C and was brought over to, say, Visual Basic .NET or C#.
PL/I will convert nearly any data type to nearly any other, e.g., something like a cast in the C family of languages, but the documentation says in very full and fine detail just how each conversion is done.
There's a relatively powerful pre-processor macro language executed by the compiler that can be used to generate a lot of source code -- can be nice.
If they have tried to upgrade PL/I, say, to 64 bit addressing, maybe more on event handling, some collection classes, a sweet DB/2 or RDBMS interface (it always had some DB syntax and semantics in the language), then it could be a nice language even now.
It's been said that the golden age of language design was the 1960s. C was designed after PL/I and was a really big step down so that it could compile on a DEC mini-computer with 8 KB of main memory while PL/I always had at least 64 KB.
On VM/CMS and MVS, PL/I didn't have access to TCP/IP, but C did, so I wrote some C code callable from PL/I that gave PL/I access to TCP/IP. Used it a little! Maybe now TCP/IP is closer to being native to the language.
I was in the group at IBM's Watson lab that did the expert system shell KnowledgeTool that was implemented as a pre-processor to PL/I. One night I stayed up until dawn and coded the logic that made rule subroutines fast, functional, etc. The key was to keep a crucial part of our run time software and the user's rule subroutine on the stack of dynamic descendancy. Gee, I got an award! The hourly rate for that night was actually relatively good!
PL/I doesn't have to be so bad!
UNIX's success in the market brought C upon us.
Actually it was a very big step down, not only in terms of what other systems programming languages of the same era were capable of, but also in terms of security and compiler research.
One just has to look at PL/I, PL/M, Algol 68, Mesa and then to C.
Just as an example, Algol for B5000 systems (1961) already had the distinction between safe and unsafe code, with unsafe program modules requiring some form of "root" access. Sounds familiar?
Early in my career I wrote some Algol. So, got to love block structured languages and adopted the style of left margin indention then called publication Algol.
I was on the computer selection committee at FedEx, and from what it had we were impressed with Algol, etc. So a Burroughs system is what we recommended and got.
Alas, it was too slow. Soon the computing was way, way behind, and Big Blue got called in and got a really big, new account. The MVT/MVS family of operating systems, with JCL, etc. was like an unanesthetized root canal procedure while undergoing a barbed wire enema, but it could get the work done.
Generally that computing is having problems with virus is a total bummer -- we should be able to run malicious code safely.
Gee, Flash keeps telling me their code is a security risk. No joke! And that's version 12 or so of their code. Flash guys, I've got more to do than be a lab rat running on a wheel going 'round and 'round downloading code.
Yup, looks like Flash guys or someone was correct: I spent all of December and part of January fighting a virus on Windows: Kept getting instances of iexpore.exe running. Yes, that is the EXE of Windows Internet Explorer (IE). So, I got those instances even without starting IE. So, I ran Microsoft's latest virus removal tools, days for each, and they found nothing. The virus was still there. So, I did a Windows System Restore back to the earliest copy I had, 2 months ago. Seems to have worked. But that was about six weeks of virus mud wrestling for no good reason.
A simple Google search shows that discussions about fighing that iexplore.exe problem is all over the Web; apparently it is a very common virus.
Flash guys, are you to blame for that one?
Flash guys: Programming lesson 1 in Coding 101: Check your input and detect any problems. If the input is not suitably clean, then refuse to use it. If the specifications of your input is too complicated for a check, at least good checks while running, then correct your specifications. Flash guys, I can understand one bug, but 12? Are you guys even trying, I mean trying to fix bugs instead of pushing users to download?
And Microsoft guys, why do you let your code let bad guys somewhere east of Moscow mess up my computer? You guys just like shipping bugs? Believe me, I do not like fixing viruses.
As I recall, Multics was written in PL/I. Prime had a version of PL/I and used it for much of their operating system.
Once Google ran a recruiting ad and I sent them a resume. I got a phone call from one of their recruiters, and his big question was "What is your favorite programming language?"
I said, "PL/I".
Wrong answer! Likely he was looking for C++.
Come on guy, I want to do something other than fight memory leaks -- in PL/I if some work is some part of my code raises an On Condition (a software version of an interrupt, which might also have been from a hardware interrupt), then in the code block that gets executed by the On Unit that has been established for that condition, can decide what to do with the interrupt. So, one thing to do is just to kill off that work that raised the interrupt. So, say that the block of the On Unit that gets executed is X in code block B, some code in block B did the function call that got the work going, and just want to kill off that work. So, from block X just do a Goto to the statement label want in block B and not in block X, and presto, bingo, that work is killed off, that is, wiped clean by the wrath of non-local Goto. So, in particular, likely all the storage (all except based storage) allocated by that worked to be killed off is freed. So, look, Ma, no memory leaks from exceptional condition handling!
Google guys: The original definition of C++ was just as a pre-processor to C. And you want to assume that that is a lot better than what IBM, George Radin, etc. did with PL/I? Do you have anyone who understands PL/I?
I got introduced to computers with a Timex 2068 back in 1986, but as a language geek I always researched the old systems.
Additionally I managed to use a few languages of the Algol family for systems programing, hence strenghting my belief in safe systems programmin.
One of the best quotes I keep recalling is the Turing Award speech from Hoare, where he mentions his company customers were against having support to disable bounds checking in Algol compilers. Yet here we are.
Every time I check the CVE list I wish UNIX had never left AT&T labs.
This parallel universe is called reality and you live in the matrix.
However, that's ignoring what's made a mainframe fast and expensive for the past several decades. Mainframes still slaughter all comers when it comes to I/O, and this is no different. When you've got up to 320 fibre links available, you're gonna have no problem keeping those processors busy.
Those special instructions such as CRC32 or SHA256 are rarely needed in scientific simulations such as Monte Carlo simulations or real-time working with Bayesian Network, what this system is aimed for. In these scenarios raw power and big on-board caches what matter and on paper the POWER8 has these, but 110 GIPS is not the best bang for bucks.
This really is a cashcow of IBM, mostly because they own the "mainframe" trademark/monopoly and everybody who wants to sell a "mainframe" is sued out of the market. Without the "mainframe" title, you are dead in this market of ultra-conservative customers (banks,airlines,etc).
They wouldn't build these if they couldn't sell them. With the io and data flow rates these thing can handle, there really aren't alterntives on the market. the debate I keep spinning through in my head is will the market stop caring about the hardware (it's all just in the cloud or grid) before the mainframe goes away or will the cloud be partially mainframe.
2. Those conservative customers want to keep their data in-house. So a cloud is probably impossible to sell here.
They can either build their own "in-house cloud" (paradox?!) or buy a mainframe. The later is simpler and proven. Cost savings of the roll-your-own-in-house-cloud are questionable.
You could try to sell an ultra-reliable (maybe distributed) system for in-house use. However, you cannot call it a "mainframe".
Nowadays you don't create a single monolithic system, really expensive too, which is capable of processing all of your transactions, as it says, like 100 cybermondays every day.
Instead, you have a distributed system, with many cheap, geographically distributed servers, each one capable of a much lower number of transactions, but still quite high today... and also you can spin up new servers as needed, or destroy servers as they are no longer needed, so you control your costs when you don't need as much instead of having a very expensive mainframe 90% underutilized 90% of the time.
IBM is full of really smart people some of whom work on really hard problems using the principles of Computer science. The z13 wasn't the result of two recent grads pivoting a Yelp for Rabbits startup and the z13 isn't for companies hoping to do a billion transactions a day. It's for those doing it.
To say that the Google 'many-servers' approach is the only valid one is to overstate the trade-offs.
At a basic level this kind of sytem gives you lots of low latency bandwidth connecting CPUs to each other and CPUs to storage. That is a useful quality to have and not neccessarily easily repoducible using network and internet connections between servers and data centers.
A mainframe is typically partitioned into multiple, smaller systems. This is what IBM calls a Logical Partition (LPAR). This partitioning is done with firmware called PR/SM. This capability has been around since the 3090 (mid to late 80's). Within an LPAR, one typically installs either z/OS or z/VM (there are other systems too, but less commonly used this way). Running in an LPAR is 'bare metal'. Within z/VM, there's typically a mix of guest types, but this is where Linux would typically be configured to run. z/VM has some pretty impressive capabilities and has been around for a LONG time. It's very stable. The biggest things that you would probably dislike about it are: (1) one uses a 3270 emulator to do day-to-day system-level admin work, (2) it only runs (legally) on real mainframe hardware, and (3) much of IBM's jargon is dated and would unfamiliar to people coming from x86.
The modern mainframe hardware and software has an impressive feature set for virtualized networking (networking is all software defined and runs within the box). This means you can set up hundreds or thousands of Linux guests and have them on networks that are all virtualized within z/VM. And of course, z/VM guests can be created and spun-up on demand and stopped on-demand. This has been there for decades.
IBM now has support for OpenStack in z/VM. From what I've perused, it seems to be quite slick. Assuming it all works as advertised, this would make folks coming from x86 feel much more at home.
Many seem to think that distributed systems are so much cheaper. You can't just think of it as the price of the rack server though. You have to include everything in the mix to understand the total cost: hardware, software, networking, people (headcount and consulting), power and cooling, floorspace, and intangibles (such as capabilities). When you do the math, mainframes running Linux under z/VM are often fairly economical (YMMV).
Once you throw off the stereotypes and objectively dig into the modern mainframe, you might be surprised at what you find.
Because the other OSes which run on the hardware are so obscure that non-technies would never have heard of them. I'm morally certain CMS and z/OS (or whatever MVS is called these days) and VM are still in use, but Linux has the penguin and the media coverage. Therefore, it gets the inches.
If you want the best with insane price level, buy into Power/AIX. If you are not really sure, please don't. You need to train administrators (with larger setups) - or prepare to shell in a lot of money, the hardware contains nasty licensing surprises, the hardware is rather expensive, and so on. When you get it running, it's insanely powerful indeed, and the platform itself nearly never screws up (if you tested properly for Monday parts), and the virtualization is insanely good in practice.
Some potential customers have bought Power/AIX, some not. The ones that did are still years after customers, and to my knowledge pretty happy. I believe IBM is actually gaining customers. Slowly, but they are. Those that made the choice knew what they got into, and are not going to change their views.
This is my experience as well. IMO, Power/AIX has the most sophisticated virtualization available. IBM just has no idea how to market it.
I was talking to an ibmer a while ago and AIX was mentioned. We agreed the platform had technical merits, and on a pure platform level was superior to x32/x64 but the network effects of commodity hardware had already won the platform game. No one writes software for any other platforms but intel/amd64. AIX lost for the battle, similar to the reasons betamax lost to VHS.
Transaction Database --> copy --> analytics db = slow
In-memory transactions with built-in analytics = fast
My bank recently alerted me because I swiped my credit card at one gas station away from home but did not buy gas (the pump was out of gas) and then swiped again at another gas station in a 3mi radius within 30 minutes (obviously since I needed gas).
Their system picked up these two transactions, figured out they were both gas stations, geo-located them to be out of my area, calculated the distance to be close enough to reason it was just one person (and not my wife using her card elsewhere), and triggered an alert because the first transaction was really just a $1 pre-auth with no actual charge.
While none of this is really that complex if you built a fraud-detection system from ground up with these requirements, imagine running this rule on a few hundred million transactions per day. Now add in a thousand more fraud-detection rules, scoring algorithms, and pattern recognition for uncommon usage. Then hook it up to the mobile app to further reduce chances of fraudulent usage by tracking user's location, businesses silently checked-in, and alert preferences. IBM is targeting companies that need this.
Or you know, any kind of transaction because there is no difference.
Fuck I hate marketing.
So I am wondering why they use TPS as a unit when it can be anything.
And of course, anybody considering buying such a machine will spend quite a bit on evaluating how it performs under their load, just like SPEC (https://www.spec.org/) says something, but not all.
Also can you elaborate more on how ASM instructions can be record oriented?
http://en.wikipedia.org/wiki/IBM_801 http://en.wikipedia.org/wiki/John_Cocke
Dunno how much it actually influenced the Archimedes though.
IBM licensed the 286 for the IBM PC/AT. Acorn wanted to use the 286 for the Archimedes, but Intel would not allow licensing. This forced Acorn to create their own processor, resulting in the first RISC based consumer PC.
Perhaps that snowballs into our current tech environment.
eh.. sorta.[0].
IBM had significant stake in Intel during the time period we're talking about, and that whole period was filled with anti-trust talk, which turned into actual anti-trust fire a few years later in the 90s.
I may not have been blunt enough in my other comment, but I was alluding to the idea that IBM and Intel may have been involved in anti-competitive behavior against Acorn, which caused them to innovate in such a way that created the systems we enjoy today (modern mobile). The idea that the pressures on the company spawned innovation.
[0]: http://articles.latimes.com/1987-08-29/business/fi-1393_1_in...
CP67 was intended as a means for interactive, time-shared operating system development. So, right, could run CP67 on CP67 -- once that was done 7 levels deep.
The combination CP67/CMS was, for the time, a total dream for a time sharing system.
Stop malicious code? Sure: On CP67 write and run any code, any instructions doing anything you want with any data you want, and you just cannot bother any other users. So, could run malicious code safely.
I used CP67/CMS and PL/I from National CSS Time Sharing in Stamford, CT to schedule the fleet at FedEx. Founder, COB, CEO Fred Smith's remark about the output was "Amazing document" "Solved the most important problem facing FedEx". The Board was pleased and a nice chunk of funding was enabled. It literally saved the company. No, Fred never gave me my promised stock, that once he said would be worth $500,000. Add a few zeros for now.
Since then CP67 was called VM, and for years IBM's internal computing was done on about 3600 mainframes around the world and all running VM/CMS and connected with VNET which was a lot like the Internet except the communications were via bisync lines and the routing was done by the mainframes themselves. No, that setup didn't depend on Systems Network Architecture (SNA). Yes, there were a lot of fora!
The advantages of running on VM were too good to pass up, so eventually essentially all production IBM mainframes were running their operating systems as guests on VM on the bare metal.
In virtual machine, IBM was way out in front.
When I started learning how OS/400 work, now IBM i. I was quite interested to see the execution model of having everything stored as bytecode, with a JIT kernel doing AOT compilation at install time.
Something that is kind of being done on the Android and Windows worlds, attempted on Oberon and Inferno, but not with the extent that the OS/400 does it.
So no on that one.
Visualiation? You mean virtualization. That was a given, Turing should be credited with that, after all the whole idea of a universal Turing machine is that it can emulate any other Turing machine (including another universal one).
One reason for industry standards is that they distribute risk across the whole industry. So if NIST crypto is broken the entire industry goes down together; one company wouldn't have a competitive disadvantage.