....I am making an argument against the idea that you can have "four nines" availability out of a single server.
And yet, in the real world, that's what many single frame mainframe shops manage. Probably because a single frame isn't at all a single frame internally. The only singleton in the equation is the cabinet itself, and (but only if the operator wants) any particular operating system/middleware/application instances that the operator chooses to run that way (which would typically be development instances, not production).
No, they're doing that because somebody had a memory leak 30+ years ago (back when those might have mattered from a system stability point of view), told the operators to reboot (IPL) every week, and the operators have never updated their procedure manuals. I'm not joking. If you tap an operator on the shoulder and say, "Stop rebooting now," problem solved. (Or, for the ultra-conservatives, switch to once every month, then once every year, then not at all.)
Or, in the alternative, somebody decided 30+ years ago that they'd tell the users the system would be unavailable each weekend for a couple hours for "system maintenance" (back when that might have made sense), then, many years ago (and particularly with Parallel Sysplex), when the operators no longer needed the "maintenance window," nobody bothered to tell the users the good news and make it official. Again, I'm not joking. I was at a meeting at a Fortune 100 company where the conversation went something like this.... Web service team: "We need 24 hour Internet access, but you're down every weekend for 4 hours. We want to move off." Mainframe operator: "You need 24 hours? OK, you have 24 hours." Web service team: "But you didn't do anything." Mainframe operator: "We stopped IPL'ing that LPAR at least 10 years ago, and I'm now declaring you have a 24 hour Service Level Agreement. I'll update our SLA just after our meeting. Need anything else?" Web service team: "Uh, that'll do."
If you want 99.999+ percent business service uptime, factoring in both planned and unplanned outages, and even disaster recovery, just ask for it. (Or ask for some other level if that's too much.) It's no myth, it's the owner's choice, consistent with their budget and "reasonable" (but not superhuman) operations staff.
Emulating a $500 PC Server on a $500,000 mainframe... yeah, that sounds real cost-effective!
Then why aren't you driving a Yugo (I presume)? It has a lower price, doesn't it?
If you run this simultaneously in 1000 virtual machines, do you need 1000 Windows licenses?
That's up to Microsoft. I can't wait to see Microsoft's mainframe price list.
Most of that thousands to one virtualization is based on the same idea that is driving commodity virtualization ala ESX, most servers spend most of their time idle.
That's part of it, but it's not the only part. Otherwise we'd see thousands of virtual machines on a single ESX core, and that's just not what's happening. (The virtualization ratios per core are pretty small. Still useful, though.) Virtualization also places heavy stresses on cache, memory, and I/O performance. IBM System z10 machines are no slouches on CPU -- they have the highest clock speed (4.4 GHz) CPUs with more than 2 cores (they're quad) on the market -- but they balance that with kick-ass cache, main memory, and I/O performance. They also have hypervisors (PR/SM and z/VM) which are extremely refined and uniquely co-evolved with the hardware over decades. Add that all together and you begin to understand why the virtualization ratios get much higher in real world use.
Yes, the usual myriad of analysts (Gartner, Forrester, etc.) have looked at the issue, and, to net it out, they find both IBM System z virtualization and VMware useful and cost-effective. ("Both" is often the right answer, because some workloads do better on one or the other, and most businesses have a mix of workload types.) One bright dividing line has been that VMware obviously can virtualize Microsoft Windows directly (such as Microsoft Exchange and Microsoft SQL Server), while System z could not. But System z has been adding whole operating systems to its virtualization repertoire, including OpenSolaris for System z and now (with Mantissa) Microsoft Windows, so it's moving into new ground.
Cost isn't the only issue of course. Qualities of Service (QoS) are also important. It's undisputed that IBM System z offers the highest QoS characteristics attainable for any particular operating system(s) that run on it. (That is, if you run, say, Linux on an IBM mainframe, from a QoS point of view it'll be the best Linux possible.) High QoS requirements form another bright dividing line.
If you have, say, 1.000 virtual Windows desktops on one mainframe you will still need to employ all the staff that would be needed to run a 1.000 machine network minus the staff needed for dealing with hardware failure.
Really? Because that's not how it works for Linux on IBM mainframes, and there's plenty of experience with that. There really are fewer operations staff required, and those fewer staff deliver much higher quality service.
To keep the answer simple, a single mainframe is a cluster. Everything inside is redundant and hot swappable. And it offers those advantages whether or not the software cares to know about it, and without a single dollar of extra labor. (Although, if the operating system and middleware participate, such as z/OS and DB2 data sharing to pick an example, even more amazing things are possible.)
For example, if a CPU core fails, an IBM System z mainframe will detect it, prevent the mis-executed instruction from surfacing to the operating system, take the core offline, provision a spare core, and continue executing without interruption. All without even the operating system programmer having to write a single (probably buggy) line of code. That's just one example. There are huge differences and huge advantages for many (not all) businesses and applications.
Good point. The first comment about "unusual word sizes" was really pretty funny, because the commenter quite obviously has little understanding of computing history. It was the IBM System/360 (the ancestor to today's IBM System z mainframe) that defined the 8-bit byte and 32-bit word as industry standards, influencing CPU architectures (including Intel's) right to the present day. Otherwise we'd probably have multiples of 6 or possibly 7 bits as our foundational standard for computing. (And there was a lot of pressure during the System/360's design to cheapen up the hardware and slice off a bit or two.)
Perhaps the original commenter would like to open up a command line in Microsoft Windows Vista and count the default number of columns. That number is 80. Why 80? Because, coincidentally about 80 years ago, someone at IBM decided that tabulating cards should be 80 columns wide, and IBM's cards were more popular than Remington's. Yes, Grasshopper, Microsoft Windows has an "unusual" column width that persists to this day.
A rock store eventually closed down; they were taking too much for granite.