What can you accomplish with an IBM mainframe that you could not do on x86 Linux servers? How is the premium in price justified?
For the type of organizations that run workloads on IBM mainframes, there are three drivers: * Legacy: The application was written for the mainframe, cannot run on anything else, too expensive in terms of dev and test time to re-platform * Business value: This is the big one; these workloads make their companies 100s of millions to billions of dollars per year. The price premium for running this on a mainframe is a rounding error. * Reliability: With the cloud, I hold the opinion that the average x86 application is less available/reliable than a well-run pre-cloud application (which already included HA, etc.). Mainframe apps and hardware blow all this out of the water.
FWIW: I programmed mainframes briefly early in my career, I am quite familiar with the ecosystem.
Every calculation in the CPU is replicated. If it shows any sign of failure it will try to migrate threads off the failing CPU to other CPU.
DRAM is RAIDed.
There is a disaster recovery capability that can replicate several data centers within a 70 km range via optic fiber. If one of them burns, get flooded or hit with a nuke the others will pick up the slack automatically.
Not at the OS level. Back when I was doing penetration testing nearly every organization that had IBM mainframes would suffer pretty severe outages just from our basic scans and doing things like checking open ports. They were also super duper easy to break into 90% of the time.
Also, most of the software running on mainframes has been running for decades. Which means they had like 40+ years to work out all the bugs. I'm 100% certain that if you took any given "modern" software stack (take your pick!) and very carefully applied patches to it for 40 years without ever adding any major new features it would be equally as reliable.
If your stuff isn't secure how could you possibly make claims that it's reliable? If I said, "resilient" (synonym for "reliable") instead would it make more sense (in the scope of IT stuff)?
Depends (like pretty much everything else) on your definitions, I guess.
I could see defining "security" and "reliability" along orthogonal axes:
"Sure, it's a bit leaky, but it's very, very reliable -- including reliably leaky."
Confidentiality Integrity Availability.
The reliability is indeed legendary for the usecases it is originally designed for (running Z/OS or TPF, IBM DB2 on Z/OS, CICS and COBOL batch jobs). However, IBM marketing folks will try to sell you on specialty processors that can run Java applications, Linux VMs(s390 arch) etc - that’s where the reliability rails come off.
Most serious mainframe users I worked with had their legacy applications which did one of the original usecases I mentioned, and it runs reliably. At the hardware level, redundancies are engineered at every level. You can hit plug CPUs like blades on a running system for maintenance, and replace them. Same with memory modules, storage devices etc. Upgrades to z/OS are also so thoroughly documented that you can avoid downtimes or plan for minimal ones..
https://en.wikipedia.org/wiki/Jazelle
The point is that Java or Linux workloads could be run on some other CPU and face competition but this is not the case for z/OS. Thus there is a reason to lower the price for workloads that could be easily migrated but keep it high for captives.
Back in the day all the essential software, e.g. flight control, which still largely runs on S/360 today, had to be proven to be correct. There are mathematical concepts/processes that allow that.
The firmware is proven to work correctly. The compiler is proven to work correctly. The OS is proven .... I guess you get where this goes.
The problem is: Nobody today even wants to learn "how to prove software" anymore. I tried to teach a class at my University a few years back, and 12 out of 12 students dropped out in the first 3 lessons. My usual dropout rate is close to 0%.
For non-essential software you are probably right. After 60 years of development all the bugs are gone; Plus the developers know their machines (hardware & software) inside out. No chasing after ever changing platforms, standards, APIs & SDKs.
I would probably word that as what can a mainframe do that commodity servers can not. They are more reliable and allow for spinning up a vast number of workloads with incredibly fast virtual networking. If one needed a "cloud" of Linux nodes, a Z16 would be insanely fast deployment and teardown of workloads on demand and those nodes could talk to each other with very little latency and high throughput. Some companies and organizations need a small private cloud (think pod or blast-radius contained mini region) that can spin up their cloud region on demand. Even in a catastrophic failure, the fault blast radius is limited to the group of racks comprising the mainframe.
The cost is not just hardware. Anything IBM is going to be IBM supported and your business would have factored that into the ROI/TCO. The contracts are very expensive but you have the highest level engineers a phone call away and if you require it they can also remotely diagnose problems and have replacement parts in {n} hours based on the contract. Not everyone needs this of course which understandably leads many to question the cost.
It's the same thing that Linus thought wasn't important at all, ignoring the advice of his teacher, Andrew Tannenbaum who was trying to teach him about microkernels, and why they were better[3].
In the PC world, we don't even have ECC memory in most desktops, etc. That's why Rowhammer[1] and all the other things like it work, because it's substandard RAM that we've grown to accept in the name of cost savings.
We keep finding Ersatz versions of Capabilities, first in Virtualization such as VMware, then in Docker and containers, and now in WASM/WASI. We'll eventually learn the lesson, likely in 5-10 years now. If we can keep WASM from getting corrupted, it might make it.
Sure, mainframes aren't cost competitive, because the cost of computing is no where near as important as the business it makes possible.
[1] https://en.wikipedia.org/wiki/Row_hammer
[2] https://en.wikipedia.org/wiki/Capability-based_security
[3] https://en.wikipedia.org/wiki/Tanenbaum%E2%80%93Torvalds_deb...
You get what you pay for. All your new fangled software and hardware just reinvents the wheel, is buggy, is layered on an onion stack with a lifespan of maybe 3 years and takes a lot more humans and money to run, diagnose and maintain. And in three years, a new CTO will come from Wharton and tell you you’re a moron for not using microservices 2.0 and spend $100M rearchitecting it.
In the meantime, Z/Power just keep going and 4 or 5 guys run the whole show. They sleep at night and have a life.
It all ads a lot of overhead and isn't particularly reliable - frankly, I think it would make mor sense to run it on a mainframe, and have simple application code because the fault tolerance is on the hardware/OS layer.
If you are a big tech company, then the massive scale means the cost advantage of x86 is worth dealing with node failures and it's not like any single mainframe for handle it, but for a lot of internal applications that are too big for any single x86 machine, but don't require thousands of nodes, I can totally see where it makes sense.
If mainframes were actually competitive with modern server hardware everyone would be using them. Even IBM uses regular Intel hardware in their cloud stuff!
Mainframes aren't even fast... IBM will make all sorts of BS claims about memory speed and the bandwidth of their interconnects and whatnot but all of it is 100% artificial benchmark nonsense that doesn't translate into real-world usefulness because nobody is rewriting their mainframe shit to take advantage of it.
I don't know about you but in the time I've been in IT "hardware failures" that actually had any sort of serious impact on operations were few and far between. The whole point of modern solutions (everything from N-tier architecture to containers and on-demand compute/function stuff) is to make the hardware irrelevant. At my work we had a whole data center taken down as part of a planned test and I doubt that any end users even noticed (and it was down for hours because they screwed something up when bringing things back online hehe). I think something like 6,000 servers and some large amount of networking equipment were complete powered down? I don't know the specifics (and probably shouldn't give them out anyway).
The whole point of mainframes is to serve a function: All the hardware is redundant/super robust (within itself). That function is mostly meaningless in today's IT infrastructure world.
Nearly all other producers, before Ford model T, tried business models "the fastest", "the most reliable and comfortable", they are also good models, but they are just much less popular than "good enough".
This was successful big distributed corporation, which begin, when mainframes was only viable technology for them, and all their business was inside machine.
They grow decades, but once happen problem - mainframe reach it's limits, and vendor was not agile enough, and said "we will transfer all your infrastructure to new hardware, but in few months" (they understand this as will stop business for long time).
I don't know all exact details, all I know, other company made them offer, claimed they will transfer to cloud, without stopping work.
And sure, big share of their business software, intensively use very specific features of their mainframes.
And as a last trick, usually such companies don't disclosure real prices, but when you buy hard+soft+consulting in one, they offer huge discounts.
So you sure, could pay their consultants, to write on Linux, but it will cost more.
I could add, I've seen x86 server from IBM, but with IBM chipset, it was capable for more cpus than any other competitor of it's time, and for some clients this was killer feature.
___
[1]: Edit -- Removed superfluous "haven't read TFA" disclaimer.