Sun Grid Utility Goes Live for Employees

museumpeace writes "CNET is reporting that Sun Microsystems turned on its Grid computing utility, hosting large ERP applications for its employees to test out the server infrastructure and user acceptance of the Computing-as-metered-utility model. General availability is scheduled for October. The rates? "Sun is offering processing and storage in a pay-as-you-go arrangement of $1 per CPU per hour, delivered via an Internet connection". Sun is still retooling its Thin Client interfaces and support SW. Experts quoted in the article wonder if Sun can make any money this way." Slashdot also covered the original announcement back in February.

Comments

Everything Old is New Again

[Anonymous Coward]

This is just the reincarnation of the mainframe era. Everyone (Sun, MicroSoft, et al), want to put us back in the days where the storage/cpu and most importantly the applications themselves are in their "capable" hands.

I'm not even going to enertain the idea of having MY data stored on another (microsoft/sun/etc)server, and paying for the rights to access/modify it.

There is a reason it's called the PC,and not a dumb terminal.

Re:Everything Old is New Again

[Alwin Henseler]

This is just the reincarnation of the mainframe era.

And what is wrong with mainframes? Putting all the computing power in one place, is about sharing that equipment between all users, making optimal use of the hardware. As opposed to everyone having their own box that comes with everything and the kitchen sink, but sits around doing nothing 98% of the time. Talk about waste...

I'm not even going to enertain the idea of having MY data stored on another (microsoft/sun/etc)server, and paying for the rights to access/modify it.

Agreed... but I can think of many compute-intensive tasks, where the privacy implications aren't that important: scientific research, running simulations, design/engineering, etc. Besides: if that 3rd party (on purpose) leaks or abuses confidential data that you trust with them, then you know who to sue for damages. And think of what running any software/OS with an automatic update mechanism enabled, means here. Kind of the same: to a certain degree, you simply trust that 3rd party.

There is a reason it's called the PC,and not a dumb terminal.

In my view PC's haven't gotten all that much smarter in the past decades, seeing how ordinary users struggle with them every day. And not faster either, only very efficient at wasting CPU cycles on eyecandy and countless software layers. Don't get me wrong here: you can do lots of wonderful things with PC's today that weren't possible 10 or 20 years ago, and I wouldn't want to swap mine with a vintage machine for doing everyday work. But just FYI: for example a meager 3.5 MHz. Sinclair ZX Spectrum boots up in what, 1 second? Doesn't need updates, doesn't have virusses, and even if it's the first time you've ever used a computer, you can learn how to program and run a "Hello World!" thingie on it in under 5 minutes. Back in the 80's, kids (myself included) used to do this with machines that were on display in shops. Now try those things with your brandnew 3+ Ghz. machine.

Back to that mainframe issue: I can see lots of people who would love an appliance-like PC, with software on it remotely managed. Think: by your local ISP, by a user group of your choosing (like a Linux distro), you name it. With your private data stored in any location of your choosing, locally or on some other remote system.

Just putting computing power remotely as well, doesn't make much sense to me for most scenario's. The speed at which computing power becomes cheap for the masses, makes that a niche market for selected applications. As always: pick the right tool for the job.

Re:Everything Old is New Again

[serialdogma]

Should that not be +1?

Re:$1/CPU/hour?

[Neil Blender]

If your using 100% of the processor 24/7 for a full year, then no, it's not worth it. However if you need a little here and there, it can be worth it.

Not for big problems, then

[Space cowboy]

Considering that a "CPU" can be had for $400 (2.8GHz Celeron D without even trying, just a search on google).

So 24 hours a day, $400 -> 16 days work. Let's add in 25% for "stuff" (electricity costs, etc., being generous...) and you're still saying that a problem that takes 20 days or more, you're better off buying a throw-away PC and running Linux on it.

So, it must be aimed at the smaller problems. Like what ?

Simon

Re:Not for big problems, then

[Keeper]

Actually, I would imagine this would be more useful for solving larger problems that are run infrequently, where you want do do a distributd task once a month that takes 1000 machine hours and get back a result in 1 hour.

Re:Not for big problems, then

[Paul Jakma]

For the same $400 you could get 100 CPUs for four hours. If your problem divides up reasonably well, then instead of spending your $400 CPU and waiting 16 days, you could instead get your answer in hours.

Maybe you could do it cheaper by buying your own CPUs, but you could be waiting two weeks for your Dells to arrive. How much is it worth to you to get an answer within hours or days versus a few weeks or months of waiting?

--paulj

Re:Not for big problems, then

[Glonoinha]

If you have a problem that takes 400 CPU hours to run, your answer is either inanely worthless, or mind-bendingly valuable (I needed to throw one of those in there for the SETI group, but I won't say which.)
Well that or you need to optimize your code, or get a faster machine.

That said, it probably isn't worthwhile to the guy with a $400 problem - more likely they are looking to appeal to the kinds of guys that want to crack 128-bit encrypted data streams in real-time, or run two neural networks against each other in a zillion games of chess in order to teach (evolve) their neural network, or crunch two terabytes of data picked up by an Indy race team over three days at the track. Brute forcing 1024-bit encryption is totally possible, but the data isn't generally valuable a thousand years after you start decrypting it. Throw enough horsepower to decrypt 1024-bit RSA in real-time and you will find yourself rich (or dead.)

Knowing the winning numbers to the lottery thirty minutes after they are announced is pretty worthless.
Knowing the winning numbers to the lottery thirty minutes before they are picked is worth a hundred million dollars.
Amazing difference having the answers an hour earlier makes - I'm not saying that these computers will give you that much of an advantage, but I'm still saying ... I currently work on problems where an hour difference in processing time can make a single data-crunching run cost about an additional $100,000.

Re:Not for big problems, then

[Paul Jakma]

I don't believe any of them cost $100 for every 20 days runtime! As for 250%, Oh boy! I have a bridge to sell you!

You're not paying for and hence you do not have:

- data centre floor space with:
    - heavy duty UPS
    - generator backup
    - climate control
    - security
    - redundant networking
    - multiply redundant storage
    - tape backup silos / HSM
- The 24x7 staff to:
    - monitor security
    - test the generator weekly
    - monitor the backup processes
    - monitor and maintain the network
    - monitor and maintain the hardware

etc. etc. If you think your costs as "Joe Bloggs the guy who runs a few Linux PCs at home" are comparable to a corporate affair then you're simply kidding yourself, particularly when you're not billing yourself for your own time ;), and your SLA with yourself is pretty flexible and forgiving ;).

A lot of corporates have thought what you thought "Ah sure, it can't be expensive to run a few servers in our own 'data centre'", and they typically either under-estimate the costs, or they end-up with very shoddy server facilities. Then they'll have reliability problems due to:

- servers overheating cause they're stuffed into a cupboard (seen this)
- lack of staff expertise (all too common)
- utilities failures (they couldn't afford the large UPS + diesel generators + cut-over switches + electricians expenses)
- the gradual increasing burden of maintaing installed plant, which if not planned professionally slowly but surely turns into a huge sprawl of unmarked cables, till it gets to point even simple rewiring tasks are a massive (and error-prone) undertaking.

Eventually, to a lot of these types of small corporations, locating in a managed data-centre and letting someone else take care of the details becomes very very attractive. (particularly for coporates whose primary business is *not* computing).

You are almost certainly underestimating the costs.

--paulj

Flip Flop

[SilentReallySilentUs]

Sun's main problem is they are not able to stick to one strategy for rescuing themselves from the mess. Last year Java Desktop System was all over the press. 100$ per developer..I don't hear of it anymore - now it is 1$ per cpu. They are going to have a hard time getting the trust of enterprises to use a CPU in their servers. Good luck Sun, more importantly Good luck Java

Re:$1/CPU/hour?

[pclminion]

Isn't that many times what it's worth? 365*24=over $7K for a year's worth of computing. Hmmmmm...

But say you wanted to run the job ten times faster. You'd split it across ten CPUs. Each CPU would perform 1/10th the work, but in parallel, so the job gets done in 1/10th the time. But the total number of CPU-hours you've used remains the same. So you pay the same price but get the job done ten times faster.

If you wanted to do that yourself, you'd have to buy 10 CPUs and once the job was done you'd have a bunch of CPUs you didn't need.

for example

[taniwha]

think of chip design - you want about a gazillion machines for QA simulation in the 2 months prior to tapeout and they're going to be idle for the rest if the year

Mind you the cost of chip design software is the limiting factor here, not the cost of hardware to run it on

big margins

[rnd()]

Sun has some fairly substantial profit margins built in at the $1/CPU/hour price.

Consider that if you have a moderately large data set that you need to crunch it's not at all uncommon for it to take 3 hours on a 300 node cluster. That's $1800 if each machine is a dual proc machine.

So suppose Sun has a 300 node cluster, for example, where each machine cost $1800. Ever 3 hours one of the machines is paid for. In other words, a 300 node cluster is paid for in 38 days. Well, the hardware is, anyway.

I really don't know who the main clients would be of this kind of service, however. I'm guessing that if your company can't afford a 10-20 node cluster (fairly cheap) and still needs to do large scale computing, renting CPU cycles from Sun would make sense, though it would very quickly cost more than the 10-20 node cluster would have. So it's really going to benefit customers who need large scale number crunching results more quickly than they can obtain them simply by building a smaller cluster and waiting for the results, or customers whose problems involve data sets that are large enough that they need to be distributed over 100+ machines in order to be solved.

Who has large data sets like that and no cluster access? Not university researchers, not government agencies, and probably not most firms doing significant number crunching.

So I see the niche as firms with large data sets and someone who can write the MPI code, but who lack the willingness or finances to invest in a cluster of their own.

In a year or two when the same service is selling for $0.25 per cpu hour it will be a much more compelling proposition.

Re:big margins

[Com2Kid]

In response, if you read one of the articles linked to from the main article:

Sun Chief Operating Officer Jonathan Schwartz submitted a project to the Sun Grid--graphically rendering data from a protein folding experiment. It took only a few seconds, but cost $12. The 12 hours of CPU time for which Schwartz was billed was consumed by hundreds of machines simultaneously clicking away at the rendering problem for a few seconds each.

Results n seconds.

Doing a quick read through I did not see exactly how large the cluster was, but there is no reason that Sun could not scale it to a few thousand nodes, providing results in a time frame that would be cost prohibitive for most companies to setup clusters to compute.

Come on, if you are running a huge simulation a few tmes a year, maintaining a 1000+ node cluster year round is just not cost efficient.

Re:$1/CPU/hour?

[Madd Scientist]

that is why sun is trying to build a system that they can't afford. a super system for the people who need a super system to share. and sun is exactly the people i would trust to build it.

Re:$1/CPU/hour is damn expensive...

[Glonoinha]

You forgot the most expensive (and often overlooked) part of infrastructure : the infrastructure staff.

Add a few $65,000 / year staffers in there to install / support those $2,500 machines and you are looking at $13,500 per year (every year) per machine. I know, that's what my company bills my department for each server I have on the network.

Re:$1/CPU/hour?

[LurkerXXX]

We've gone over this before. The price isn't bad. You don't buy time on this system when you have one CPU's worth of stuff to compute and don't need it for a years time.

You buy time on it when you need a LOT of CPUs worth of stuff done NOW.

Imagine you have some projection software package that you need to run once a quarter for your company. You need the data within a week of the beginning of the quarter. You require 10,000 CPU hours to get the numbers all crunched. It's the only "big-computing" job you have.

On one computer the task would take you a little over a years time (8544 hours in a year). That won't quite be up to the task, remember you need the job done in a week. That's 10,000 CPU hours to fit into 168 hours of real time. You'd neeed 60 processors chugging away for those 168 hours to get it done.

How much is a 60 CPU cluster going to cost you to build? It's not insanely expensive, but it's not cheap. It looks a lot better to you to build that cluster than to spend $40,000 a year though! Right?

Wait. Clusters take up space. A 70 CPU cluster (better add in a few for redundency since this job has to be done in time) is not going to fit in the broom closet. That floor space is going to cost you.

Hmm, those 60 CPUs throw off a lot of heat when they run. Better add some more cooling to the building. Another decent expense.

Damn, look at that electric bill from the extra 70 CPUs and cooling for them. This nickel and dime stuff is starting to add up.

And now for the killer. You've got a new 70-CPU cluster. Your going to need someone to manage it. Cluster work is a bit different from what's what your used to, and your IT staff is already busy with their current workloads. It's time to hire a guy to manage the cluster. BZZZZZZZZZT. That hire alone makes the $40,000 a year for grid CPU time a deal.

Work the numbers yourself. It's not really a bad deal if you only occassionally need massive computing.

Applications that make sense

[acaspis]

As expected most comments about "who's gonna pay for this" and "it's cheaper to run your own server".

But think about business models where the grid provider sells not only CPU cycles, but also trust.

Scalable web hosting: Your PHP code is replicated on-demand to as many grid server as needed to handle your peak loads. The grid provider guarantees that server-side code and data remains confidential. End of the slashdot effect as we know it.

MMORPG: Small startups can deploy worldwide networks of game servers in no time and compete with the big boys.

The next Google: Anyone with a smarter search algorithm can go online without investing in huge datacenter first.

Rendering farm: Your CG movie is due to premiere next month, and your 10,000-node rendering farm can complete the job in time. Wouldn't you pay extra $$ to anyone who can save the day and guarantee that screenshots won't be leaked to the Net ?

Combine this with a micropayment infrastructure where the grid provider sends bills to end-users on behalf of the service provider. Huge potential.

AC

yay for slowlaris

[Yonder Way]

Not all CPU hours are the same. An hour on a moderately fast SPARC processor is not as valuable as a moderately fast Intel Xeon or AMD Opteron or a PowerPC.

Flexibility

[tez_h]

Others have noted that the cost of a bespoke system is lower. These slighly ignore the fact that this is aggregated over some time period, like a few months or a year, where the for most part of that time the computer may be sitting idle if there aren't jobs to fill up the time. But Sun's solution offers granular computing time for any number of CPUs at a linear rate, meaning you don't pay for the time you don't use, and you can pay for lots of CPUs without penalty. I would imagine this convenience and the additional conveniences of infrastructure, maintanence, security, updates, etc are what you're paying $1/CPU/hour for.

Since Sun is attepting to utilise economies of scale, as this gets more popular, presumably prices would become more and more accessible.

-Tez

I can shoot down one of these

[RelliK]

Rendering farm: Your CG movie is due to premiere next month, and your 10,000-node rendering farm can complete the job in time. Wouldn't you pay extra $$ to anyone who can save the day and guarantee that screenshots won't be leaked to the Net ?

And how the fuck are you going to transfer *hundreds of gigabytes* of data required to render a frame over the internet? How are you going to receive the data back? (2MB - 12MB per layer per frame).

Does that thing even have Renderman installed? (at $5k/CPU I highly doubt it). Does it have Shake? Does it have Houdini? Does it have Maya?

Besides that, how the fuck are you going to get approval to send _anything_ out of the studio? You obviously have never worked in the industry.

I'm also skeptical as to whether there is any use for this. What sort of environment do they run it on? Solaris/SPARC? Solaris/x86? Linux? Windows? What sort of software does it have installed? Would it ever be possible to replicate the in-house environment on this "grid"? (you know, with all the custom software, directory structure, environment variables, aliases, etc.) I know for a fact that there is no way we could outsource our rendering to Sun even if we tried.

The whole "CPU-hour" thing is a very nebulous concept. Environments differ wildly from one company to another, so you can never have a universal "CPU grid" in the same sense as you can have an electric grid.

Re:$1/CPU/hour?

[Saanvik]

Actually I think this might be very appealing to research groups at universities.

As part of your grant proposal you include a flat cost for computer time rather than costing out hardware purchases. Not only that, but you can also start your project as soon as the money is approved, you don't have to go through all the hoops to buy, ship, house, and administer the hardware.