Slashdot Mirror
IBM Wants CPU Time To Be A Metered Utility
kwertii writes "IBM CEO Samuel J. Palmisano announced a sweeping new business strategy yesterday, pledging $10,000,000,000 towards redefining computing as a metered utility. Corporate customers would have on-demand access to supercomputer-level resources, and would pay only for time actually used. The $10 billion is slated for acquisitions and research to put the supporting infrastructure in place. Will this model revolutionize the way companies compute, or is this plan doomed to be another PCjr?"
It will be a revolution until Linux becomes mainstream on the desktop and every computer on the corporate LAN is part of a cluster, when users log off the computer re-joins the cluster. Companies should look at what they already have before shelling out more money.
If the huge false start that was Application Service Providers showed anything it is that corporate customers don't trust computing resources that are outside their control. It doesn't matter if IBM can provide a better service or a more reliable one, it just doesn't feel that way - and the IT guys will never report favorably on something that will put them out of a job.
It's PCjr, it's Gavilan, it's all kind of failures. And $10,000,000,000!
See, at that time ubiquitous networking was not a way of life. Also, software engineering was not as mature as it is now WRT to virtual machines, encapsulation, OO design, etc.
Of course, all those technologies did exist then, but they can be counted on to be everywhere now. The reason mainframe timesharing gave way to PC's is because PC's could provide a more flexible and convenient sandbox to compute in, rather than the cumbersome interface of working with the mainframe in the company basement.
These days returning to the idea of computing power as a fluid resource is a good idea because the landscape has changed and the world might actually be better prepared to accept the tradeoffs since the tradeoffs are much less significant now.
don't ever let this happen. The car design scenario creeped me out. I work for an automotive supplier, and we ALREADY have to wait in line to use test equipment, testing chambers, etc. I can only imagine the local supercomputing hub monopolizing the speedy machine, creating more lines to wait in, and IBM bringing its supercomputer prices out of reach for anyone but their own subsidiaries to purchase. Could be a disaster, indeed.
hi, I like pancakes -.-- -.-- --..
The last 35 years development in computers were precisely to move away from the "metered service" model which made IBM's fortune.
On will recall that IBM's data-processing customers since the 1920's were charged by units of information stored/processed by the way of forcing customers to buy Hollerith (punch) cards solely from IBM, and run them in rented machines whose rental price was directly proportionnal to the throughtput of those (a card reader that processed 600 cards per minutes cost twice as much as one that processed 300, yet the only difference was the size of the pulley off the main motor - and you could upgrade by having an IBM tech that came and changed the pulley for a bigger one).
So is it that the ghost of Thomas Watson sr has made a comeback to IBM's board of directors????
No, it's an Application Service Provider, the next step in outsourcing. The idea wasn't all that popular during the dotcom craze; is it any better now?
...phil
"For a list of the ways which technology has failed to improve our quality of life, press 3."
I don't know what everyone is comlaining about. Take a look at the wonderful user-oriented, monopolistic services companies like the Phone and Cable companies currently provide (Qwest, Verizon, Cox, etc). This is a terrific model to emulate. And think of all the lovely intelligible Taxes the government could add to your monthly computing bill.
0xfeedface
The problem is that most reasonably sized departments need an IT staff anyway. Having them run a mail server or the like isn't that big a deal. While some things can be effecient to subcontract out (i.e. your web server) often it is easier to have it on sight.
There are exceptions. But I think that only a few IT functions really can reasonably be marketed out. I think IBM's marketing strategy will work - but only for a small niche.
What IBM has said is that it hasn't got anything new to report but that its still here. If you look at their figures $10Bn works out at 3.5bm for the consultancy firm they purchased, a few billion for Grid computing, and I guess a couple of billion for linux. With a bit of spare change for research.
Why are they doing this? My guess is that CFO's keep complaining about the cost of computing resources. A multinational with 10,000 desktops still has to ask for clusters and supercomputers for serious work while TFlops of processing are sitting idle on the secretarys desktops. Hard Disks, which used to be able to just about hold the OS, Office suite and files now have 10's of GBs of wasted storage.
If you're serious about using computers you want to use resources efficently. And from IBM's perspective so how does this idea sound ...
IBM sells computers to a firm, it then sells the software to turn all their hard disks into a P2P file storage system so that you never lose that important document ever again. Instead of a new cluster - set all the desktops to process data overnight as a massivly distributed system. (using IBM software), installed by IBM engineers under the direction of their new consultants. And of course the only real option for this is Linux.
A single, nice, neat package. A single point of contact and massive economies of scale. Now assume that their contract allows them to use/sell spare cycles and their revenue stream suddenly improves a lot.
The revolution will be in revenue.
Currently IBM big customers buy a new machine every four years or so, they pay a yearly maintenance bill. IBM has trouble predicting it's revenue quarter to quarter, in a downturn everyone stops capital expenditure and IBM mainframe sales plummet.
Under this model everyone should pay less but they'll pay every month like clockwork.
Computer Associates has a similar scheme for software. You rent your software on a monthly basis.
On a technical level I'm all for it. I have a suite in my current site that is run yearly and takes for ever. Currently IBM has a big box sitting here and we just sip from it, until year end when we max it out for like two weeks. Let me rent time on a huge box and I'll be happy. Gurantee my data and response time and I'll be ecstatic.
Bad news: the rates charged per byte/cycle/whatever ought to drop by 50% every generation (12-18 months these days).
More bad news: typical supercomputer code is usually bummed (at least a little) for the particular hardware it runs on, to get the last factor of two or so for performance. If you rent crunchons, can you afford to rent generic crunchons and give up that last bit of optimization?
Good news: if you can get around the bad news above, this could turn supercomputing into a lease-vs-buy situation, and when the computer you buy essentially depreciates 50% every generation, leasing might be a win.
To a Lisp hacker, XML is S-expressions in drag.
I know people who have generators or windmills and are connected to the electrical grid. When power demands are high, the power company actually pays THEM for their surplus power.
If I have a nice Linux cluster that meets the "standards" for the grid (whatever they are), can I sell cycles back to the provider? Or is it just one way, in which case I'm trapped into doing whatever the grid wants me to do.
This would (did) work in the early days of computing, when it was virtually unheard of for anyone except for large fortune 500 companies and the US gov't to have access to computing power.
Why would anyone be tempted to return back to this model? How many sub $500 or even sub $200 dollar computers, will it take for IBM to realize computing power isn't rare or expensive?
And if a company or organizaion needs incredibly massive computing power is needed then can turn to companies like this to provide the solution, again using cheap generic pcs.
To some it all up this is stupid, and now Palmisano looks like another idiotic buzzword chanting CEO. This will be yet another blow to IBM, and it will soon (IMHO) join the growing stable of companies (Compaq, HP and the "new" Cisco) that have been screwed by a clueless greedy CEOs. Somebody needs to cancel his subscription to Business 2.0
As long as they keep it within reasonabl bounds.
People do not want "shared computing"; they do not want to put their data on "borrowed computers" nor do everything on "rented computing power" or "rented space". IBM should realize that most people will still want their applications and most of their processes and files on their own computers.
What IBM should be offering -- and what it seems like they're offering -- is loaning supercomputer time to people (for a price) for specific tasks which they can't accomplish in a reasonable amount of time on their own computers. This is a reasonable and useful idea; however, it is hardly new at all. At the University of Rochester, there are shared computers within biology labs, where people dump some heavy-duty computing operations and pick them up later. This went on during the 60's when computers were so expensive no-one could afford them. In short, this is hardly new nor revolutionary, though IBM may be putting a new twist on it by trying to use it as a business model.
It makes sense. After all, most people don't need supercomputing power for the majority of their tasks; why spend money on a supercomputer when it'll be unutilized 90% of the time? But what IBM can do is maximize supercomputer utilization by selling a percentage of its resources to various customers; these customers save money because they pay on a per-need basis.
For example, I often run Bayesian phylogenies. Recently, I ran a Bayesian phylogeny with about 50 taxa in it. This took 7 days on a dual G4 (2x 800MHz) Mac. That's with all of the computer's power focusing just on that. The time requires to complete the trees increases at a steep rate as one adds more taxa. If I was doing 200 taxa, it would have taken two or three months.
So this can offer a great service to many people.
social sciences can never use experience to verify their statemen
The real world has a huge diversity of applications - most enterprise applications can't just be outsourced for maintenance, ongoing development and so-forth, unless by the people who developed it in the first place. Exodus and the many colocation facilities of the late 90s and early 00s wanted to offer services sort of like this, but it just doesn't work - they don't have the talent in shop to do it, and can't learn everybody else's apps.
If by "IT department" you mean IBM will operate databases, Apache web servers and J2EE app servers and other commodity applications in their own datacenters, then I do believe it, but again that is what a lot of high-service colos were doing several years back (many of whom went under). The economies of scale aren't there - the only people who would think they are are those who think of "IT" as some mythical blob of computer operators, and don't realize the mix of trained sysadmins, developers, and so-on that make up "IT".
And the ASP model - well, the problem there is that though the company that developed an app is well suited to actually host and operate the app, if a corporation adopts that model, then their apps will be hosted and operated all over hell and high water. I mean, this is fundamentally the web services model, and it's nice for a lot of things, but I don't think anybody believes it is going to make corporate IT departments go away and allow the centralization of all computing work into a big IBM datacenter. I'll believe that when I see it.
Yes this concept is Timesharing on Steroids, but check what this CEO guy has already done- sold the commodity hard-drive biz and gone for Linux in a big way. He is clearly not risk-averse and assuming we all agree Linux is A Good Thing (and certainly a way to beat on Sun and Microsoft) he is not stupid. So what is he doing here?
Posters who are focusing on the U-word (utility) need to see that IBM doesn't want Joe Citizen using this. The profit levels for dealing with the general public just aren't there for IBM- Big Blue is all about the corporate or government cash.
In a word, cost savings for premier customers, i.e. the kind of people who will run up huge MIPS but not on a constant daily basis. Scenarios that come to mind beyond the car engineering ones are banks/companies/bureaucracies who have monster End Of Month/End Of year processing but reduced needs otherwise, websites that have a lower average use threshold except when the Super Bowl commercial airs, and disaster recovery (keep your disks mirrored offsite, if a disaster occurs call IBM, get your virtual mainframe up and switch to the offsite array).
With IBM's sysplexing and workload algorithms in play, tying in 'outside' 'puters will waste few resources.
I suspect that IBM's ultimate goal is disk farms on user sites and CPUs at IBM's Grid Ranch. With the CPUs under IBM's care they can really drop the TCO for the machines themselves.
That reminds me, the real cost of operating mainframes nowadays beyond the staff is the third-party licenses for the support software- security, tape libraries, etc. That's because traditionally the software vendors license by MIPS on the machine, not MIPS actually used in your LPAR (logical partition, a carved out virtual machine on a mainframe). Whenever you increase the MIPS of your machine, the third-party vendors will bleed you dry (which ultimately loses IBM customers as they go to cheaper alternatives).
IBM is beating on these vendors by competing in their arena to drive TCO down, and is also trying to get them to meter their actual usage under z/OS. So this grid thing is just a logical extension of what they are trying to do to not get run over by Moore's Law and the cost of running The Big Box.
________________________________________ History Must Not Fall Into The Wrong Hands ___________________________________
Maybe it's just that I'm reading at +4, but I'm suprised that nobody's mentioned Multics yet. The first thing that popped into my mind when I read "computing at a public utility" was Multics. I mean, the whole point of the system was to make computing a metetered utility. Not that any significant conclusions can be drawn from this, since Multics' failure had nothing to do with the business model, but more to do with them having overly ambitious goals for the project.
my sig's at the bottom of the page.
A business will always choose a fixed cost over a variable cost. But there's many points of view.
From a system administrator's point of view:
I work in the data processing indstury, and we have a 12-way NUMA box as our mainframe. We moved from a 16-way SE-70 that we'd had for seven years earlier this year, and our software has already expanded to max out the capacity of the new NUMA unit - to the point where we've upgraded it several times.
We'd continue to expand if the perception was that we have unlimited resources.
From a business point of view:
Even if we could do our dp activities on someone else's mainframe, we would still have our system administration costs for systems that can't be moved out of the building, so our costs don't go down. We would also have to maintain in-house development machines, because we wouldn't want to pay someone else for the endless compiles that we would need while developing new software.
Additionally, we already have a huge, unamortized investment in fixed dp assets.
Currently, our systems process for 24 hours per day to meet our needs. If we were to do these same activities on a metered system, we would probably not have to process as long, but if costs are over $5,000 per month metered, it's not worth it, especially since there are no cost savings except for the cost of amoritization of our main hardware.
Corporations buy unmetered data lines because they don't want to have to deal with variable (and, in case of a slashdotting, extremely high and exteremely unstable) costs. Trying to sell a service that has a variable cost structure is good for a company, but buying a service that has a variable cost structure is bad for a company. The only time buying becomes good is when the company can't provide it for itself, as with electrical power and telecom. But it's easy to buy/build your own mainframe-class computer for less than $10,000.
--
Vote for your hopes, not for your fears - Vote Third Party