"A base box with hardly no MIPS"?
You sure do have a purty mouth....
The IBM mainframe may not have a lot of direct competition when it comes to z/OS-based batch and transactional work, but the story is different when it comes to Linux. There's plenty of competition among Linux platforms, and Big Blue can't ignore the pressure that Moore's Law brings to bear. So IBM is slashing the prices it …
...you already have one. (or really, not less than 2, as buying 1 frame is almost never done...)
If you already HAVE a few frames, because you need to run many DB2 databases, have lots of legacy code, or whatever. Lets say you've licensed 4,000 mips on each, and have lots of spare capacity. Now, adding some IFLs to run Linux is not only easy, it;s cheap, and adds to no other infrastructure.
But lets go with your base assumptions, a completely new mainframe chassis (2 actually) vs enough VMWare nahalem chassis not just to run the guests, but to also add VMotion and enough additional redundant chassis and SAN storage to handle full failover of a datcenter. (If you have enough guests to warent a mainframe, you probably already have more than 1 datacenter in your building, and if you don;t, it;s a REAL good idea you look at your infrastructure redundancy design...)
Lets also look at a few other details your cost analysis missed:
Storage: In the mainframe, hundreds of Linux "guests" can use a single binary image. Hundreds of VM ware guests needs hundreds of times the base OS image storage. Yea, you need data storage, but that's the same price either way, but a 5GB OS image on the mainframe compared to 0.5TB on 100 linux guests is a HUGE difference. (since the base frame has all the I/O components and dedicated I/O processors you'd need, you're just basically buying DAS storage, but for the VM Ware systems, it;s SAN for you...)
DR: (goes to storage as well). On VMWare, your 100 guests need significant storage either on tape or disk for backups. On the mainframe, just a base copy is all you need to back up, plus the differrencing data. There's also no licnecing needed to do this on a mainframe, but each of your 100 linux guests in VM ware needs a backup agent and so does the VM Ware host.
Software (honestly, this is the biggest one!): Lets look at Websphere, Oracle, or any number of other apps you need to buy for these linux guests. On the Host side, 1 "processor" license covers the entire IFL. 5 websphere licences (at 120PVUs each) covers all 5 engines, and hundreds of Linux guests. In VMWare, each of those processors is 120PVUs (if you're talking nahalem xeons). A single rack chassis costs almost as much to license as the whole mainframe... Add to this MQ, DB2, antivirus, scanning and remediation clients, monitoring clients, and more that the mainframe basically either takes care of already, or simply doesn't need).
VMWare itself: Ouch thats's expensive... how many cores did you license it for on the small boxes? including enterprise support, HA, vMotion, storage replication (you do plan on having more than 1 SAN chassis, right, and are not putting 10-20TB of mission critical system images on one bot that could take 3-4 days to recover with even the best DR solutions, right?), management software, and more?
Switching/routing: Mainframe, just needs a few 10Gs to connect to the network. Multiple VMWare chassis? lots and lots of cabling, inband and out of band VLan configurations, fully redundant rack to rack cabling and switching, not to mention the HBAs, SAN fiber switches, and more...
memory: Yes, mainframe memory is nearly $2,000 per GB. Of course, the base frame comes with some significant memory (128GB i believe), and each IFL adds to that. But again, single binary imaging allows a very small memory footprint change per additional guest. We're adding between 64 and 128MB of active memory footprint per additional guest. on VMWare you're likely adding 1-2GB if not more. Sure, you can get 2GB kits dirt cheap, but with only 8 or 12 memory slots, that's not getting you far. Likely you'll be buying 8G and 16G kits, which do not scale linearly in price so well... and the frame has a good head start....
We've been deploying zVM for about 3 years. Upper management CONSTANTLY looks at each new project to compare costs. We actually look at everything I metioned above and more to make our determineation to deploy new or upgraded systems on zVM, VMWare, BlaceCenters, or standalone boxes. If the software runs on zVM (keeping in mind it's not linux, but Linux for OS390X hardware platform, which is a seperately compiled version of Linux, hence the no-need-for-antivirus position), it is ALLWAYS cheaper than VMWare (usually by half or more), and in some cases we've seen a 10 fold savings (especially when dealing with per core or per guest licence models).
Another alternative could be a 8 socket x86 server with 512GB RAM. For instance the SUN X4600M2. That machine can virtualize at least 82 servers, including Windows servers.
And holds several world performance records:
Mainframe can only virtualize Linux servers, not Windows servers. That is not optimal, from a virtualization view.
Also, if that x86 machine uses 2.8GHz quad cores, then we have 32 cores, each running at 2.8GHz. This means that one 8 socket x86 server has GHz equivalent to 8 sockets x 4 cores x 3GHz = 89.6GHz = 90GHz.
As we know that 1 IBM MIPS == 4MHz x86 (according to this Linux expert):
We see that 90GHz x86 == 22.000 Mainframe MIPS.
And one SUN 8 socket x86 server starts at $25000. Quite cost effective solution to get 22.000 Mainframe mips on a x86 server. And you dont need special expensive software. Just use your normal software.
Another thing is that when IBM states that a Mainframe can consolidate thousands of x86 servers, IBM quitely hides that the x86 servers idles at a few percent work load, whereas the Mainframe runs close to 100% load. If the x86 servers started to do some work, the Mainframe could never virtualize that many x86 servers.
As usual the x86 view is to simply focuses on the cost of the tin and get all excited about speeds and feeds whilst ignoring that most inconvenient of truths - HW accounts for less than 20% of the TOTAL cost of ownership. The largest cost driver in IT today is SW and people. Even if a z10 BC 5 way could run a 5th of the servers of the equivalent cost of Intel boxes, it is doing it on 5 CPU's instead of 78 cores.
The SW license savings for IBM SW or even Oracle software on a virtualised z will far exceed the aquisition cost of the HW - usually saving multiple millions in SW licensing and support costs over a typical 4 year view - enough to buy 4 z10's if necessary (a z10 can scale to 10 IFL's)
Oh - and you don't have to buy those IFL's again when you upgrade - they move across to the next generations for free.
Oh - and an existing System z typically already has an excellent DR solution in situ so if you have one you get DR for free for your Linux workload (nice :-)
Oh - and if you need to add capacity it is non-disruptive - pretty important when you are running 100's of production instances (
Oh - how do you add capacity on Intel ? buy another box, rack, network switch, powerboard, add a project manager and a 6 week lead time (all for free of course)
Oh - and an IBM System z and z/VM has 99.999% up time
Have I forgotten anything ? Ah yes - a System z has up to 128 co-processors handling I/O, something that comes in mighty handy when you have a 100 virtualised operating systems making continuous I/O calls to a SAN. In Intel ALL I/O is handled by those cores - thats why you need so many. And thats why the SW is so expensive on Intel - you are paying SW licenses for cores that are handling disk I/O overheads instead of executing real application workload.
Thats why most of IBM's System z customers are deploying Linux on z - because it performs, is highly available AND is cheaper and easier to manage than the toasters that filling up the other parts of the datacentre.
"....HW accounts for less than 20% of the TOTAL cost of ownership...."
Yes maybe that is true for Windows. But how much does a decent Mainframe cost? 10 million USD? Compare that to a few x86 servers. With x86 servers you can get more performance, for one percent of the price of a Mainframe.
If we talk about lowend Mainframes, then any 8 socket x86 server smokes the Mainframe. Not to mention AS/400
Yes, I suspect a Mainframe does the job very well. But I am saying that not everyone can afford 10 million. It is much cheaper to get 10 x86 servers and handle 90% of a Mainframe's workload. And if we talk about CPU performance, the Mainframe lags far behind x86 servers.