This was some great read
replies(1):
This was a great article - thanks for sharing!
the article here ends around 1971 --- the mainframe would later save IBM again, twice, once when they replaced aluminum with copper in interconnects, and then when some crazy IBM Fellow had a team port Linux to s390. Which marked the birth of "enterprise Linux", i.e. Linux running the data centre, for real.
they also moved on three more CPU generations since that redbook, to z17.
I think it's Linux on Z that makes it sexy and keeps it young, in addition to a number of crazy features, like a hypervisor that can share CPUs between tenants, and a hardware that support live migration of running processes between sites (via fibre optic interconnect) and the option to hot swap any parts on a running machine.
It's doing a number of things in hardware and hypervisor that need lots of brain power to emulate on commodity hardware.
_and_ it's designed for throughput, from grounds up.
Depending on your workload there may be very good economical reasons to consider a mainframe instead of a number of rack-frames.
They feel fantastic when running Linux, but, if you don't need all the reliability features that come with the platform, commodity hardware might be a better choice for the kind of workload that has evolved on Linux.
> Depending on your workload there may be very good economical reasons to consider a mainframe instead of a number of rack-frames.
Absolutely - it makes a lot of the administrative toil disappear. I know clusters are sexy, but getting the job done is always better.
But most people don’t want to deal with the hassle of dealing with IBM.
In this case they will just use a mainframe, even it isn't cheaper in the long run.
For legacy companies yes but it would be very hard for new YC companies or existing non-mainframe companies to create a spreadsheet showing how buying a new IBM Z mainframe would cost less than the latest commodity x86 or ARM servers in the cloud or on-premise.
The IBM pricing for mainframes makes sense for legacy companies like banks and airlines with a million lines of old COBOL code that want to keep it all running with the latest chip technology. (The mainframes from a few years ago are coming off the lease and they need to upgrade to whatever new mainframe IBM is offering and sign another lease & service contract.) So, IBM mainframe prices are very expensive -- but legacy companies will pay it because migrating the code away from mainframes can be even more expensive.
It's similar to expensive enterprise pricing of Oracle, Embarcadero Delphi, Broadcom etc that takes advantage of existing customers already locked into their products. Virtually no new tech startup with a greenfield project is going to pay Delphi $1599-per-seat for each of their developers. Only existing customers stuck with their investment in Delphi code are going to pay that. (https://www.embarcadero.com/app-development-tools-store/delp...)
But some companies do endure the costs of migration to get out from IBM lock-in. There are lots of case studies of companies shifting their workload from mainframes to AWS/GCP/Azure. I can't think of a notable company that did the reverse. Even a mainframe hardware vendor like Fujitsu quit making mainframes and shifted to x86 running an emulation of their mainframe os.
Yes, IBM mainframe can run other workloads besides COBOL like Java and C/C++ but no company that's not already using mainframes would buy & deploy to IBM's Z hardware for that purpose.
Set to some nice Philip Glass music to boot.
Did it though? Or was it the gradual phasing out of mainframe-class hardware in favour of PC-compatible servers and the death of commercial Unices?
Proprietary Unix is still around. Solaris, HP-UX and AIX still make money for their owners and there are lots of places running those on brand-new metal. You are right, however, that Linux displaced most of the proprietary Unixes, as well as Windows and whatever was left of the minicomputer business that wasn't first killed by the unixes. I'm not sure when exactly people started talking about "Enterprise Linux".
It depends. As we have seen the other day, HPE has a machine with more than 1024 logical cores, and they have machines available to order that can grow up to 16 sockets and 960 cores on a single image of up to 32TB of RAM. Their Superdome Flex goes up to 896 cores and 48TB of RAM.
I believe IBM's POWER line also has machines with more memory and more processing power, but, of course, that's not the whole story with mainframes. You count CPUs that run application code, but there are loads of other computers in there doing a lot of heavy-lifting so that the CPUs can keep running application code at 100% capacity with zero performance impact.
> It's the easiest, most reliable way to scale a classical transactional workload.
And that's where they really excel. Nobody is going to buy a z17 to do weather models or AI training.
https://juliusgamanyi.com/2018/12/28/wardley-maps-an-illustr...
https://juliusgamanyi.com/2019/06/18/wardley-maps-an-illustr...
https://juliusgamanyi.com/2020/06/25/wardley-maps-illustrate...
Now, reading the article, this "[rivalry].. So intense was it that sometimes it seemed to exceed the rivalry with external competitors.” reminded me of something old about Motorola.. where rivalry went to the need to reverse-engineering other depts' chips via 3rd party acquiring them..
www.chicagomag.com/Chicago-Magazine/September-2014/What-Happened-to-Motorola/
Perhaps this is just the attitude that drives Mr. Kay's point home - do individuals who are interested in CS have little value for who and what has come before them?
BUT. I worked at a place that used IBM 360s. We ran stuff for engineers, a lot of Fortran along with assembly code. We had so much stuff going on we could not code up and run things fast enough. The engineer/scientist got frustrated.
Then one day an engineer brought in an Apple II from home and ran the programs on that.
The earth shook. The very ground beneath us moved. Tectonic plates shifted. The world was never the same again! I think it was Visicalc.
Later there were other things. Soul Of A New Machine. The Mac.
I wonder how the compute power of a current high end smart phone compares with and IBM 360? I know the graphics chip is better.
Which is the Mainframe vs commodity server dichotomy.
An IBM 360/20 on the small side, however, ranged from 4kB to 32kB which was similar to home computers circa 1980, before it is routine to have a complete 64kB address space.
Where the 360 crushed home computers was in mass storage, 9-track tapes could store 80MB contrasted to floppy disks that stored less than 200kB. Large storage compared to memory meant a lot of focus on external memory algorithms, also there was already a culture of data processing on punched cards that translated to the 360 (e.g. terminals have 80 columns because punched cards had 80 columns)
https://en.wikipedia.org/wiki/IBM_System/360#Basic_hardware_...
This may be true, but because there's basically no on ramp to running on a mainframe, there's no way anybody is going to try it unless they're already on a mainframe. Or maybe unless they really need something that only a mainframe can provide. But most companies can live with some downtime, and once you can live with some downtime, you have options for migration between sites, and options for migrating loads so you can swap parts on a stopped machine. Splurging on network infrastructure with multi-chasis redundancy is an easier step to take to get to a more reliable system than building against a totally different system architecture.
Cornell had one of very few IBM 3090s with a vector unit (to compete with the Cray) just before I showed up, but when I did IBM had donated a message-passing based supercomputer based on the Power PC architecture. I only saw a 3090 (no vector unit) at New Hampshire Insurance which I got to use as a Computer Explorer.
(2) I was taught in grad school in the 1990s to use floats if at all possible to reduce the memory requirements of scientific codes if not actually speed up the computation. (In the 1990s floats were twice as fast as doubles on most architectures but not the x86). I really enjoyed taking a course on numerics from Saul Teukolsky, what stood out in the class as opposed my reading to the Numerical Recipes book which he was a co-author of, was the part about the numerical stability of discretizing and integrating partial differential equations. If you did it wrong, unphysical artifacts of the discretization would wreck your calculation. Depending on how you did things rounding errors can be made better or worse, Foreman Action's Numerical Methods that Work and later Real Computing Made Real reveal techniques for managing these errors that let you accomplish a lot with floats and some would point out that going to doubles doesn't win you that much slack to do things wrong.
Oracle's big reason for doing so was because they could charge more for Oracle on Linux, and still get to a lower total cost of ownership than Oracle on Solaris.
Oracle began this in 1998. By 2006 they had their own Oracle Enterprise Linux distribution.
The 360 was introduced in 1964. The 370 was introduced in 1970. The 3033 was introduced in 1977. The Apple 2 was introduced in 1977. So, yeah, if you were still using 360s contemporary with an Apple 2, no wonder the engineers were frustrated.
When I was growing up (decades ago, mind), my dad kept trying to convince me to get into machining. Lathes, mills, tool and die making, etc. In his line of work, he saw lots of machining companies basically paying all education and training expenses for new hires because their experienced old timers were retiring faster than they could be replaced. (He also thought computers were a fad...) I'm sure it was a good opportunity at the time. But nearly all of those companies eventually closed up because offshore manual labor was and still is way cheaper. If I'd had taken his advice, I'm pretty sure I would have had to reboot into another career at least once by this point.
Steam tech is much, much more interesting than it appears at first.
Not that there's a lot of historical context to things as far as which people did what - most of that sort live on in names of techniques and methods (Rankine cycle, de Laval turbine, Carnot efficiency, etc.)
Well...sure, you could put bigger storage on a mainframe. It's just money, after all. But you could put a tape drive on a home computer. And bigger disks. And a card reader, for that matter. Where the 360 really crushed the home computer was in aggregate bandwidth, via the Channel architecture. An Apple 2 could just about keep up with a floppy and a display. A 360 could keep up with dozens to hundreds of tapes, disks, card readers, terminals, printers and other things all at the same time.
Large storage compared to memory meant a lot of focus on external memory algorithms
I would agree with that. I would just argue the real mainframe advantage is a whole-system one and not point to a single factor (memory size).
I do believe that is true.
IBM did have the Model 90 series, which was on the way to being a supercomputer.
Bur during that time CDC with the 6400/6600 etc was likely bigger in scientific computing.
A current high end smartphone has around 10 billion transistors.
From https://gunkies.org/wiki/IBM_System/360, IBM made 11-12 million SLT modules per year in the late 1960s, with less before that. Each individual SLT module contained a handful of transistors. Therefore, in transistor count alone, a single smartphone has more transistors than IBM produced through the 1960s. And this is before we consider the fact that clock speeds today are much higher than they were in the 1960s.
Your smartphone literally has enough hardware to outcompute the entire world circa 1970.
Isn't it amazing what over 50 years of Moore's Law can do?
> as I move into my 4th and likely last decade of working with computers
Don't despair--I am in my sixth and still program every day.
You can get a partition of IBM Z to run in the cloud. The cost is about $5/hr for the smallest configuration.
> Splurging on network infrastructure with multi-chasis redundancy is an easier step to take to get to a more reliable system than building against a totally different system architecture.
Yes and no. If you truly need that redundancy than the mainframe is going to provide a much better version of it. SYSPLEXs and LPARs are some insanely powerful technologies.
They announced it and started getting a huge number of orders; IBM hired 1000 people a month and kept up that pace of hiring for 5 years.
Still, 55 years of doubling transistors at a given cost every 2 years is about a 190 million fold transistor difference for a given cost. Clock speeds have improved by a factor of 1000 on top of that. Even with performance tradeoffs for battery life in a smartphone, there is no surprise that the phone should have more compute power than the world did in 1970.
http://www.roylongbottom.org.uk/Cray%201%20Supercomputer%20P...
I agree with this, and all your other reasons why mainframes will remain legacy platforms.
But man, how crazy is it that an industry that is 100% productivity driven and (until AI data centers, at least) labor-cost-constrained would snub its nose at paying a fraction of a percent of those developers’ salaries to put good tools in their hands.
Another outlier was Deep Blue. Estimated cost of about $10 million. Its estimated strength was matched by top end PCs about 9 years later. Your phone today is better than those PCs.
Progress is relentless.
The 6502 in the Apple could address 64k of RAM. Any class of problem requiring more memory would need a real machine.
As far as a personal machine with comparable capability, RISC brought that to the market with the first MIPS R2000 in 1985.
mainframe pricing usually is not by pound of iron but per cpu cycle. unless you say nah I'm buying the thing, not just cycles. soyou have a choice of cloud pricing or hardware pricing.
for significant pure Linux workloads that need an own on premise cloud the TCO pricing is surprisingly competitive with any home grown solution.
even if you don't _need_ an own on premise cloud, AWS and similar donate come for free and there may be a break even there, too.
that said, legacy host software, Delphi, cics, Oracle, DB/2, cobol etc, you named it, yes that's a different story, with all the extra software licensing.
but a pure Linux mainframe is surprisingly competitive not just in compute and throughput but also in pricing.
plus you get quite some high availability in silicon. think ECC on the ALU, hot stand by spare CPU, multipath everything, if your business depends on uptime your system architecture becomes much simpler on highly dependable hardware.
so that spreadsheet compares own people and software construction vs IBM tax in some columns.
lol I start sounding like an IBM sales person https://www.ibm.com/products/integrated-facility-for-linux
I just did Linux on Z for a while and I loved it, so please bear my enthusiasm :-D
I also appreciate that my Internet connection is about 33 million times faster than my first modem. It'd take me over a year to download what I can slurp in about 1 second now, even if I could afford the 7 thousand floppies it'd take to store it.
Progress, yo.
and that pulled all other vendors along, HP, Dell, Fujitsu, likewise for software...
and it all started with IBM officially supporting and pushing the hobbyist student project Linux on the holy Grail of enterprise compute, (of 1999/2000): s390
It's not competitive when you consider what long-time IBM mainframe customers actually do. E.g. SABRE airline and travel reservation system was one of the first IBM mainframe customers in 1960 with the IBM 7090 and then upgraded to System 360 and then the newer Z mainframes. SABRE's multi-decade experience with IBM mainframes with mission-critical business applications means they are one of the world's foremost experts on its capabilities & costs that's not biased by any IBM marketing or sales pitches.
Even with all that in-house experience, SABRE still chose to gradually migrate off of IBM mainframes to less expensive tech stacks. Some have touted IBM's TPF (Transaction Processing Facility) on the mainframe as compelling technology but that still didn't dissuade SABRE in ~2001 when they migrated the airfare pricing application to Tandem (Compaq/HP) NonStop servers running UNIX. (https://www.computerworld.com/article/1339621/has-mainframe-...)
They then started a 10+ year effort to migrate more mainframe workloads to Google Cloud. E.g. from https://www.sabre.com/insights/a-journey-to-tackle-legacy-co... :
> We moved 93% of compute capacity from physical data centers to the public cloud, resulting in a 50% decrease in our compute costs.
> We migrated more than 120 million bookings from a mainframe-based system to one using Google Cloud’s Spanner database, without impacting customer operations.
JP Morgan Bank is another example of migrating from IBM mainframes to cloud. If anyone out there truly thinks that running a new greenfield Linux workload will be be cheaper or even cost-competitive on a new IBM Z mainframe, just pause and consider if you truly know something about IBM mainframes' OpEx that multi-decade IBM customers like SABRE and JP Morgan don't already know.
Why would any new customers in 2025 willingly buy into IBM Z mainframes if they see existing customers spending billions trying trying to move off of them?
I suppose the argument could be made that the 68000 was first, as both it and MIPS ended up in gaming consoles (Sega Genesis vs. Sony PS2 and Nintendo 64).
However, MIPS eventually scaled to 64-bit, was well-known and heavily exploited in supercomputing applications, and was used to produce the film Jurassic Park. The 68000 had a far dimmer future.
Yes, the x86 line did supplant them all, but only with AMD's help. Had Itanium been Intel's final answer, MIPS might be much stronger today.
Mainframes had much more powerful distributed IO for multiple hard drives, tape systems and such.
But the CPUs were really not all that powerful at all.
Many people wildly underestimate just how insanely fast and power-efficient today's commodity computing is compared to high-end supercomputing from earlier decades.
A Raspberry Pi destroys any supercomputer earlier than around 1990, and gives some later models a run for their money.
An Apple S7 watch processor is maybe 25X faster than a Pi 5.
And so on.
4.3BSD was basically designed without any bespoke hardware platform of its own. They commandeered DEC’s big iron, for the most part, to “dual boot” before dual-booting was cool. 386BSD began to enable PC-compatible hardware and really cost-effective “server farms” even before Linus was a twinkle in Finland’s eye.
Moreover, various BSD flavors were empowering admins to breathe new life into legacy hardware, sidestepping and bypassing the proprietary software channels. Linux, on the gripping hand, remained x86-unportable for awhile after BSD (and Xfree86) was running everywhere.
Personally I ran Minix-286 at home; at university we enjoyed a “recreational” VAX11 running not VMS but 4.3BSD.
Flash forward to 1998: from the arid but air-conditioned Sonoran Desert I received gently-loved twinned Apollo 425t systems with memory upgrades; I installed OpenBSD on both, as well as my 486DX100! It was a homogeneous OS environment with heterogenous hardware... and the 486, with Adaptec SCSI & a VLB #9GXE64 Pro, could boot into Windows 98 and run the Cygwin X server, or DOOM or Quake.
There was a golden age when a dude could walk into any surplus yard, grab Big Iron Unix Boxes, take them home and bootstrap NetBSD. On anything. Bonus: BSD originated in USA/Canada, for a trustworthy chain of trust. (Oh Lord, the encryption export technicalities...)
Linus played follow-the-leader for years.
Carnot, Thompson, Clausius, Gibbs, Rankine, Boltzman etc all made big historical contributions to modern understanding of Thermodynamics.
And for Fluid Dynamics: Euler, Bernoulli, Mach, Stokes and so on.
And if you are looking for someone more modern I'd say Ergun (Packed bed's, Fluidized Bed Reactors etc).
All builds upon "steam science"
Nearly every installation site and their internal support director agrees that the z/OS installations are far less expensive to support, simply because the minimum level of experience for a z/OS engineer is about a magnitude more than a cloud engineer.
-- https://en.wikipedia.org/wiki/History_of_Linux
It was all about Solaris, AIX and HP-UX for .com businesses.
I was there, our Linux deployements were only used internally, we never sold those workloads as UNIX targets on our product.
About then was the IBM 360/91 with a 60 nanosecond cycle time. So, that would be a clock speed of
1/ ( (60)/(10^9) )
= (10^9) / 60
= 10^7 (100 / 60 )
= 10^7 (10 / 6 )
= 10^7 (1.66)
1.66 (10,000,000)
16.66 (1,000,000)
= 17 Mhz
4000/17 = 235
(16)(4000) / 17 = ~4000
Edit: apparently these cartridges actually are 18TB, 45TB is "compressed" capacity. Still though, terrible marketing aside it's still factor of 225,000.
you don't have racks with blades but one box or two, that you upgrade every few years and other than that your hardware side is covered.
rewiring networks, reattaching storage etc all remotely, CLI, and most importantly you only pay what you use.
you need one guy or a consultancy for the initial etc up and for fancy Config changes but once it's essentially set up you simply have a large Linux data centre and do Linux stuff.
oh and you can hot plug additional physical CPUs into a running VM and pull it out and reassign when it's not needed, likewise RAM. that's somewhat difficult with commodity hardware.
and there _are_ companies that value this and choose Linux on Z as a pure Linux deployment.
but you have a point. if nobody in the building knows this, nobody's gonna call IBM and ask for a quote. it's simply not on anybody's radar.
The IBM 360 is ^4 RAM and the iPhone is ^8 RAM. So the iPhone is ~10,000x more powerful.
Then there's CMS which runs under VM; theoretically it could run as an LPAR, but I've not seen that as running DIAGNOSE (VM system call) to the LPAR hypervisor could be disruptive. It would be interesting to drop DIAGNOSE into a z/OS environment, but I suspect it would be intercepted.
I much prefer z/VM to z/OS as a development environment. At one shop, I developed products under VM for deployment under both VM and MVS.
Many early S/360 installations ran 7070 and, I guess, 1401 emulators.
Eventually with VM, the SIE (Start Interpretive Execution) instruction appeared. A form is running LPARs.
Early processors were typically 1 or lower. Modern stuff is all superscalar piplined and out of order and can do way more than you’d expect. Not to mention SIMD operations and other technologies. Branch prediction is probably better on the new chips too.
And with more RAM and cache algorithms can be chosen with different tradeoffs for less instructions.
16 cores at 4ghz was a thing like a decade ago - chips today might have the same specs but are definitely far faster.