Sun Enters Grid-Computing Rental Market

mOoZik writes "BBC News is reporting that Sun Microsystems has launched a pay-as-you-go service which will allow customers requiring huge computing power to rent it by the hour. "Why build your own grid when you can use ours for a buck an hour?" asks Sun's COO Jonathan Schwartz."

Comments

$1 per CPU hour

[Bryan Ischo]

Note that the cost is $1 per CPU hour. This means that if your application uses 1,000 CPUs, it will cost you $1,000 per hour. Since the target applications are large problems that are not easy to solve without huge CPU resources, the cost for most applications will be quite a lot.

And yet, it will probably be very cost effective for certain applications, where the cost of building and maintaining your own computing grid would be prohibitive.

Somehow the thought of the world moving back towards "mainframe" style computing with truly "central processors" and everyone with a terminal in their home is comforting in a nostalgic sort of way.

Re:$1 per CPU hour

[carpe_noctem]

Damn, I was just about to host a 1,000 person q3a tourney for a cool 50$. =/

Re:$1 per CPU hour

[Wiz]

I can see where this is going.....

1. Charge $1 per hour of CPU time on your cluster.
2. Lower the speed of your processors.
3. Runtime of tasks increase. So your $1 does less.
4. PROFIT!!!!

Cool, no ???? step! :)

Re:$1 per CPU hour

[macklin01]

Indeed!

Our research group recently bought a small cluster (around 40 processors), and as the project moved forward, it found that finding a good place to put it with sufficient cooling and power infrastructure was quite a bit more costly than originally assumed.

The idea of renting a lot of computing power without bothering with these issues is very attractive. -- Paul

Re:$1 per CPU hour

[Uber Banker]

Haha, how cynical! But given they're in competition and not collusion with IBM (for now) plus the 'traditional solutions', perhaps that is second order in their minds.

Where this is going of course, as the article touches on, is the commoditisation of raw computing power, making it a product like iron ore, coal or oil. Genericism will, IMHO, be a really interesting force behind evolution of computational techniques over the next 10 years.

With genericism perhaps there can be no monopoly in provision of computing power, or more specifically for Sun, there is genericism until we get to the buts-and-bolts Solaris and then we get a monopoly in proven scalable OSes and as long as they can hold off Linux snapping at their heels under the reigns of Linux, then they'll have their monopoly and that is step 4; which leaves holding off Linix presumably as step 3, just you forgot to denote it '???'.

Re:$1 per CPU hour

[RazorJ_2000]

Valve, Steam, and Blizzard should get together and just switch their barely-reliable systems over to this.

Re:$1 per CPU hour

[nine-times]

Note that the cost is $1 per CPU hour.

Do you know that because you actually RTFA, or did you just look at the banner ads that have been running on /. since before the story was posted?

(Anyone else find it funny that the news posting on /. announcing this seemed to come after it was already being advertised on /.?)

Re:$1 per CPU hour

[Anonymous Coward]

All of the software required to be written in Java. End of discussion.

Re:$1 per CPU hour

[AviLazar]

Valve/Steam just host the pointers to the servers - so the barely-reliable is based on whose system you are accessing. It is similar to blizzard - companies pay blizzard to license the software and give them a pointer.

As for the grandparent post - you can host your UT3 tourney - just charge each player $3/hour :)

Re:$1 per CPU hour

[crunk]

I'm not sure exactly how this will be implemented, but there is one thing that concerns me about the "terminal" approach to computing.

You are now much more dependent on your connection to the mainframe. If you had an Internet/ISDN/whatever connection to a mainframe server, and that link went down, your employees don't have anything to do.

Here is an example. At my work there is much communication/work completed via e-mail. All sites for my company had their own MS Exchange servers for e-mail. If our Internet link went out our local e-mail worked just fine (although, we obviously couldn't send anything over the Internet).

However, once our company was bought out, our new masters implemented a centralized e-mail system. Can you guess what happened? *ding ding ding* You're right! Router went down and all 100+ employees at our site were worthless.

Maybe one day Internet connectivity will be dependable enough for this to work. I just don't think we are there yet.

It would seem to me that having local application servers and dumb terminals _would_ work.

Re:$1 per CPU hour

[Bryan Ischo]

If you had RTFA yourself, you would have known the answer to your questtion.

Re:$1 per CPU hour

[phyruxus]

A few years ago when AMD was still an upstart and beowulf clusters weren't mentioned in every other troll post, I wondered if there might be money in something like this. I thought it would be cool if AMD set up a warehouse in some cold place like Alaska and threw a load of overclocked commodity chips down, and used all the natural freezing air to get free and efficient cooling.

I don't know what my point is.

Re:$1 per CPU hour

[o1d5ch001]

But I can't work now with the Intern[NO CARRIER]

Re:$1 per CPU hour

[evilviper]

Somehow the thought of the world moving back towards "mainframe" style computing with truly "central processors" and everyone with a terminal in their home is comforting in a nostalgic sort of way.

The world has been heading this way for quite some time now.

VNC has single-handedly caused a huge come-back in remote computing. Projects like the LTSP and Sun's thin-terminals have helped as well. But none of those things ever seemed a candidate to replace a computer on your desk, because of performance, and price of a not-so-thin client on every desk.

What I believe will really make a huge dent is fourhead Unix systems. On a dirt-cheap PC ($200) you can attach four sets of keyboards/mice and 4 monitors, and have everyone working on it simultaneously. You don't need $100-$400 thin clients for everyone, and you don't need a full-fledged computer to act as an X server/VNC client. The only additional expense over an individual computer, is the videocards and keyboards/mice.

The thing keeping mainframes down right now, is the lack of a graphical equivalent of RS-232. If Unix systems could provide a display over Firewire/USB2, then a single mainframe could power dozens and dozens of dirt-cheap graphical terminals. Unfortunately, nobody has come forward and set any sort of a standard, and shared Ethernet really isn't up to the task.

Re:$1 per CPU hour

[owlstead]

So you bought a couch that didn't fit in your living room and now you're buying another house? Smart! Strange things some people dare to say on slashdot.

Re:$1 per CPU hour

[JamieF]

Is it time for the DAN (display area network) to appear in the enterprise? :)

I don't think Firewire or SATA have the cable length for this sort of thing. Ethernet is cheap, though, and if the economics made sense, a server chassis with a 32-port gigabit switch or something similar wouldn't be technically too hard to put together.

I agree, though; the problem is that you'd need something like a $50 or $100, very small display endpoint (featuring ports for power, 10/100 Ethernet, SVGA out, and USB) to make it work. Some little embedded Linux running an X server, or better yet, a little x86 cigarette-pack-sized gizmo net booting would be enough. Hell, the whole thing could run Windows too, except that Microsoft would much rather sell you a Windows and Office license for every endpoint.

Re:$1 per CPU hour

[SirKron]

More like ...

1. Charge $1 per hour of CPU time on your cluster.
2. Lower the speed of your processors.
3. Add spyware to the grid.
4. Runtime of tasks increase. So your $1 does less.
4. PROFIT!!!!

Re:$1 per CPU hour

[nelsonal]

I've thought about it in Montana (our power company caught the .com bug and tried to become a telecom, before they went bankrupt they laid fiber to some pretty podunk towns). I was thinking that you could run heat exchangers (even a nice passive one would work well to minimize your cooling costs in a data center. Also for the 5 cold months your air conditioners would run quite efficiently (as you are pushing heat to cold air). You would have to be careful of the eventually large ice sheet that would form under your heat exchangers, though (or you could rent it out to the local kids to play hockey on). I'm surprised no one has tried it (a few engineers would live like kings with anything aproximating coastal pay in some of the local areas).

Re:$1 per CPU hour

[evilviper]

I don't think Firewire or SATA have the cable length for this sort of thing.

SATA? That came out of nowhere...

Firewire has a max single length of 35meters IIRC, and only requires a repeater for longer distances.

Ethernet is cheap, though, and if the economics made sense

No, no. It's pretty obvious that Ethernet isn't up to the task. First of all, it requires very high-end thin-clients to handle the storm of packets, not to mention the massive load packetizing all the data causes on servers. It would really require putting router hardware into servers. Ethernet is cheap for the same reason Winmodems are cheap... Firewire is cheap as well, and it doesn't require every device to have large processors, interrupt handling, etc.

Firewire/USB2 make for a far better fit than ethernet. I personally think Firewire is the best option, but the rest of the computer industry obviously disagreed with me on that point.

Re:$1 per CPU hour

[SpaghettiPattern]

Note that the cost is $1 per CPU hour. This means that if your application uses 1,000 CPUs, it will cost you $1,000 per hour.

What a bold way to express your lack of knowledge of CPU costs. $1 per CPU hour is dead cheap (even for low end Solaris systems.) $1k for an hour of 1k CPUs is even cheaper. Do you know how much you have to invest to get this amount of CPU bandwidth? And do you realize that the odds are you will not load a 1k CPU system efficiently? Have you paid attention to your screen saver reporting the average CPU load? I have spent time doing capacity planning for a large institution and I have seen CPU prices for IBM mainframe, Solaris and Linux on Intel/AMD so I know what I'm talking about.

OTOH, the BBC article has no detail information and the buck-an-hour quote must be taken with a grain-of-salt . I also could not easily find useful details on Sun's site. But very likely, $10 per CPU is still dead cheap for commercial organisations.

Can Spamers use it

[Anonymous Coward]

Does that mean spammers don't need the grid of zombie windoze boxes? So sun is competing w/ msft.

Re:Can Spamers use it

[Anonymous Coward]

Spamers don't need it because:

1. Their grid of windows boxes is bigger than all of sun.
2. Their network is Free (and people said Free Software doesn't run on windows).

Re:Can Spamers use it

[dogfull]

no

spammers need network bandwith, not computing power, which is what this solution delivers

Re:Can Spamers use it

[freedom_india]

You have to add:

3. The grid is more robust than Sun

advert

[sashang]

haha the front page advert made the news

Re:advert

[LBArrettAnderson]

How is this offtopic? It's being advertised all over OSTG. A leaderboard advertisement i saw over an hour ago was this very thing.

Woo-hoo!

[Saint Aardvark]

Eat my SETI@Home dust!

IBM

[Anonymous Coward]

Looks interesting but IBM are doing a similar thing with their "on demand" servers and have been for a while.

Beow....wait a minute

[anakin876]

imagine a beowulf cluster of these.....oh wait it is already a grid. hmmm. Actually this could be really cool. I wonder how many companies will want to use it though. I think the security concerns (handing Sun your information, the possibility of someone else recovering the information at a later date and so on) may scare some companies off.

Re:Beow....wait a minute

[I8TheWorm]

Most companies, unless they are pharm or bio, won't really have any use for this. It's the labs that are folding DNA or running weather pattern predicitons that would have use for Sun's service.

Re:Beow....wait a minute

[Anonymous Coward]

Good for movies too. Building a rendering cluster each time you make a movie is expensive.

Re:Beow....wait a minute

[Uber Banker]

Most companies, unless they are pharm or bio, won't really have any use for this.

Financial mining and oil companies will have a HUGE use for this: mining for geology, oil for fluids and financial pricing, financials, obviously, for financial pricing (ranging from neural nets to weather).

So once we've stripped out pharma, bio, oil, mining, financials... there aren't many large company industries left! Plus, I'm really interested how academia will take to this.

Re:Beow....wait a minute

[Mindwarp]

Finance is a big user of grid technology, and also of buying in capacity like this. There are distinct peaks of requirement during the day (close of business of the different international markets) where calculation demand spikes. The rest of the day demand is limited to intra-day valuations, ad-hoc pricing etc. which is of a much lower computational volume than the big market close 're-value everything' runs.

Re:Beow....wait a minute

[Coryoth]

Good for movies too. Building a rendering cluster each time you make a movie is expensive.

Depends on how you look at it. Weta Digital spent a fortune on a render farm for LOTR, but that money got recouped on the film, and they now rent on time on the render farm to universities, research groups, and businesses. If you really need that powerful cluster now you may as well build your own and then sell off time on it at 90c per CPU hour when you're done and it's not under heavy load anymore.

Jedidiah.

Re:Beow....wait a minute

[bitslinger_42]

Lots of companies (bigger companies, that is) have a use for this. Computational Fluid Dynamics (CFD) for modeling things like exhaust flow, noise propegation in an engine compartment, etc. and Finite Element Analysis (FEA) for figuring out necessary material strength for body panels (can you guess I work at a traditional manufacturer?) are both good candidates for grid computing. There are other places doing similar rentals, such as the NCSA in Champaign. They've been doing it for a few years, and it looks to be going fairly well.

Re:Beow....wait a minute

[I8TheWorm]

Having been in the financial industry for some time, I can say most don't use neural nets. They would rather count on the expertise of the Harvard graduates they paid a ton of money to decipher the market (the best fund manager I saw at AIM, where I worked, had an undergrad from SFA though). I'm sure there are a few that use them, but they would be in the minority.

I'm sure you're right about the rest though. But there are tons of other industries outside that list. Retail, real estate, engineering (mostly CAD), municipalities (run like a business anyway), wholesalers, logistics companies, etc... That was a partial list of the companies I can think of that won't have any use for grid computing.

Re:Beow....wait a minute

[AaronGTurner]

Features such as containers in Solaris 10 allow you to address some of those security concerns, along with encryption at rest and in flight. Pharmas are likely to be unhappy about those security issues for some time to come all the same.

A practical concern, however, is the cost of shipping data to the compute. If your jobs are compute intensive rather than data intensive then this is great. If your jobs are data intensive then shipping the data for peak compute demands may not be feasible.

If the business requirement for that peak is strong, though, $1 an hour for the peak compute but with an associated transport cost may be cheaper overall than maintaining a large compute facility large enough to meet that peak compute demand adjacent to the data storage. However there is always the option for a business to sell the spare compute cycles themselves in the way Sun are suggesting.

Basically the big determinants here are how much data you have and how much you are prepared to pay and/or how long you are prepared to wait for it to be moved about and how much compute you can do on that same set of data once it has been transported. Ultimately we will see intelligent brokers that will allow companies to set policies for these and for provision companies and automatic agents driven by those policies go through a bidding process. There are projects such as ICL's ICENI looking at mechanisms for this on top of existing grid middleware.

Re:Beow....wait a minute

[TheOldFart]

What about the terabytes of texture maps and the terabytes of rendered images that must be transferred back and forth?

Re:Beow....wait a minute

[Uber Banker]

For finance, I was relating to the absolute use rather than calcs finance cos do rel to other techniques - I suppose NNs to weather covered the spectrum.

But what bit of finance are you in? It's the quanatitive vs. qualitative argument when quant is qual based but specification of the quant means a lot (as it does with mathematical techniques including NNs in general), and that is qual in the end. For PE then it is all in the business assessment, but semi-academic minds which grace Ivy Leagues and Wall Street fare poorly in terms of NVA.

Most major FI models use NNs as a tool do add colour to their model, they are extremely common in large finance houses or even just semi-quant niche houses. Not that NNs are used for predicting prices, but I'd prefer a NN/fractal model and Econ grad to add the qual. than a Harvard MBA and quants using brownian models. I've been very successful using fractals to trade vols.

Re:Beow....wait a minute

[I8TheWorm]

I was a quantitative analyst... just a programmer with number skills now. My team was actually awarded by Steve Lipper (Lipper Analytics) as the most advanced of its kind at the time (it's been a while now). Maybe things have changed considerably in the last 8 years, but the contacts that I still have aren't using any really advanced modeling. Rather, they (sadly) are counting on the "expertise" of their Wharton/Harvard MBA's to pull the trigger.

I actually tried to use more advanced stochastic modeling, but there wasn't a single economist or portfolio manager that would even look at the data. I got the feeling that they took it as job competition. Oh well.

It's nice to hear there are other companies using modeling like that to, at the very least, get a better top down picture.

Re:Beow....wait a minute

[Uber Banker]

Well, I'm the economist who has a math bent! So much is blindly assumed as you mentioed earlier.

But perhaps we're talking at cross-purposes. I agree almost all finance houses rely on the qualatitive 'economic' view point (despite price moves orthogonal to 'fundamentals' echo second, third, etc order guessing of the market), but many are using 'optimisation' methods in terms of NNs and possibly fractals or even kernels. I'm just a beginner with only 3 years under my belt but as far as I can see on the sell side its becomming common place to use this in FI portfolios (and FX), less so in equities (I'm buy side).

I'm interested in what you say about the past, in terms of quant I know MS used NNs and what are still some pretty bleeding edge techniques 10-15 years ago, even though it was with limited success.

Stochastic modelling ;) I can present an analysis of Brownian vs. fractal vols to my CIO and he'll say: "What was the 3 month performance? I think this is wrong because I had lunch with so-and-so political hasbeen with a name far surpassing track record and agree with what they said"! Then he'll start quoting Sharpe ratios on payoff which bear similarity to the lottery test.

Oh well! A lot of the old prop-desks have left their parents to form as hedge funds now, which is a good thing for the progress of quanity vs. pragmaticism, IMHO.

Re:Beow....wait a minute

[I8TheWorm]

Mostly what we did were linear stochastic models... autoregression, r^2 trends, etc, probably low level compared to what you're doing. I worked a bit on NN and applying chaos theory to some of the models, but none were accepted as they were "unproven" models. I'll call a few friends and see if anything new is going on over there.

I laughed out loud here when I read your comment on Sharpe ratios... I had almost the same conversation with our VP at the time :)

Re:Beow....wait a minute

[Uber Banker]

Well I like to dip my toes into a lot of things... we're pretty informal about proving something though, data-snooping abound, in the end its a trade-off with good model and using it soon enough to take advantage.

NNs are such a nightmare though - so much to define and specify - useful and lots of 'value added' but so far no program trade to make me a squillionaire ;)

Why did you call it quits and move on anyway? Once you're broken through as a quant there's quite a nice package as long as you keep up the chirade...

Re:Beow....wait a minute

[I8TheWorm]

Haha... charade is a good word for it....

On the mutual fund side of things, there comes a point where you can no longer advance without an advanced degree from a "prestigious" school. Around here (Houston) that's most Ivy League schools (Wharton school of business being the top), UT, and Rice. I didn't want to go to grad school, and was having more fun solving problems with code. I had been a programmer for 8 years when I started at AIM, and acidentally fell back into it there.

There was a little bit more that helped my decision. Our manager left, and I was passed up on the position (no grad degree) even though I had tenure and was running the group anyway. The new manager really didn't have a clue, and put more pressure on me to handle his tasks as well as team lead. He created Supervisor position, dangled it in front of me for 6 months, then pretended we'd never discussed my moving into the position. The very next morning, I turned in my two weeks.... even leaving a copy of my letter with Ted Bauer, then CEO (when I started it was a pretty small company).

Here's the funny part. The offered me an additional $1500/year to stay, rather than the $8000 promotion I would have recieved. Of course I declined. When I left, the calls poured in.. how do you do this/that? After a week I told them any more calls would require a contract at $100/hr with a 3 hour minimum per call. So they offered me the $8000 without the promotion. I told them no, I wanted $13000... that they should have done it a month ago. Three weeks later, they offered me the $13000, and I told them no, they should have done that three weeks earlier. They called two weeks later and offered the promotion with a $13000 raise, but I had already found a programming gig making much more... so I just told them to stop calling. It was nice to know I did add value to the group, and that AIM finally realized it.

Re:Beow....wait a minute

[Uber Banker]

That's really interesting, good to hear you yielded your power when you left instead of being shat upon! Must have been a bit of satisfaction after that diappointment.

I suppose I've got all this to look forward to, there are only 5 of us in front office (we're an in house investment manager running a pension fund) and I've started looking for something somewhere else so I can dig my teeth into what I enjoy and am good at a but more. I think I'm sensing that frustrating attitude you describe - its the CFA in my case: there are far better things to learn given 20 hours/week to than poor statistical techniques, incorrect economics and GAAP, yet somehow the CFA is a 'gold standard'.

All the best.

Re:Beow....wait a minute

[I8TheWorm]

Admittedly, I still get a warm-fuzzy when I tell the story. I do wonder what I'd be up to if I were still there though.

Have you looked into CPM (Certified Portfolio Manager) at all? It looks interesting, and at least isn't filled with as much GAAP CRAAP as the CFA.

Finally Sun does something that makes sense!

[SirFartsAlot]

As a SUN bigot I have been more and more challenged to find reasons to remain faithful. This shows some promise.

Re:Finally Sun does something that makes sense!

[InfiniteWisdom]

Don't worry... they'll screw it up soon enough.

Interesting to see the future...

[chris09876]

I wonder what will happen to this technology. It does seem like it could be useful for a number of applications (university research, for example). If you had a big problem that you spent a lot of time preparing, and then needed a bunch of processing power, this seems like the ideal solution. It certainly is cheaper than building your own giant cluster... but as the first poster pointed out, you pay per CPU per hour, not just per hour.

Re:Interesting to see the future...

[RWerp]

Why should I pay SUN, if government gives me access to Cray for free to do my research?

Re:Interesting to see the future...

[k98sven]

I wonder what will happen to this technology. It does seem like it could be useful for a number of applications (university research, for example).

Actually, I think corporate research would be much more of a market. For one, if you have an academic department doing the kind of research which requires heavy computing, then their need is probably going to be pretty constant, and you'll be better off building your own grid. And the ones who don't need that power on a day-to-day basis are usually picking up the slack on the university grids. Academia has a long and established tradition of collaboration and pooling common resources, from telescopes to particle accelerators.

Corporate research is a better target, where you might, for instance, need big computational resources for a certain project or contract, but not on a day-to-day basis.

Re:Interesting to see the future...

[ciroknight]

I dunno, this thing sounds perfect for our University (UofL). We have two "super computers", a newer Operton based one, and a much older IBM big iron server (supposedly the same kind as the one that ran Deep Blue). Both super computers are open to all computer science students as a means for centralized storage and backup, and for compiling bigger projects, but other than that, they go unused (as far as I know). I believe if they ever hit a big enough research project that required more CPU power than what we have (which is to say, not too shabby for a university, but it's no BigMac either), we'd be shit out of luck. That was before this offering, though.

Re:Interesting to see the future...

[nadadogg]

Then this is of no value to you, and you wasted bandwidth posting your reply.

Re:Interesting to see the future...

[tonyr60]

And if "government gives me access to Cray for free to do my research" then I have no doubt that you waste more than bandwidth.

Re:Interesting to see the future...

[rtv]

Academia has a long and established tradition of collaboration and pooling common resources, from telescopes to particle accelerators.

On the other hand, academia has another great tradition of finding it much easier to obtain captital equipment money than operating money, so many, many clusters get bought and built, but get little use. Many sit idle for months or years. Of course, no one ever switches them off, so they sit there eating power. Next time someone needs a cluster, the old one looks too slow, so another one gets built. Rinse and repeat. I've seen way too many hot, noisy, unused machine rooms.

I estimate that if it takes one unit of work to obtain one capital equipment dollar, then operating dollars (read: student, postdoc or engineer support) take around 5 units of work. Given that brains and labour are the key research resources, this seems bizarrely lopsided. Sigh.

Re:Interesting to see the future...

[AaronGTurner]

One of the advantages of Grid is to defray those capital costs by allowing researchers to gain easier access to existing resources. Traditionally academic HPC has either been isolated islands in various departments on various campuses augmented by a few supercomputing centres offering relatively centralised facilities. In theory Grid offers a way to make access to resources scattered across a campus seamless so those dusty old machines almost everyone has forgotten about can be used. There are challenges in making this scalable and robust and easy to use from an end user perspective but great strides are being made.

Ultimately, though, machines still have a limited lifecycle as user requirements for memory and turnaround will mean that older machines will reduce in usefullness and the power drain will mean that it is not economical or ecological to keen them running for the amount of use they are getting. Some are still more useful than it would appear as part of the problem is a lack of tools to automatically parallelise or otherwise subdivide computation algorithms into logical units suitable for a variety of different underlying architectures. Software engineers can create these parallelisations but with a variety of possible architectures available some automation of the process would be advantageous. However the relatively low level programming languages used for most scientific code makes it hard to automate this. I know Sun was looking at a higher level language for HPC and this would be a good development for future code as a higher level description of the algorithm could be more amenable to automatic processing. In addition something that is closer to the mathematics involved than C or Fortran might make the code less prone to bugs (compilers and translators permitting).

Re:Interesting to see the future...

[RWerp]

Because only privately-funded research brings progress to science?

Re:Interesting to see the future...

[danila]

you'll be better off building your own grid

You mean like if your electricity needs are constant, you build your power plant? And if you need bottled water in your office every day, you drill a well?

What happened to doing your own business and letting professionals handle the rest? It's not like Sun expects us to pay through the nose for the processing power. They'd like to cover costs and have a decent margin, but probably nothing unreasonable.

There aren't likely to be any overheads in running your problems on a remote computer. For most tasks you won't need a lot of bandwidth (i.e. not multi-terabit links), so if you need a cluster you might as well let the professionals build and administer it.

Re:Interesting to see the future...

[k98sven]

You mean like if your electricity needs are constant, you build your power plant? And if you need bottled water in your office every day, you drill a well?

In case you haven't noticed, quite a number of factories do have their own power plants. And the ones which use a lot of water (e.g. breweries, paper mills) have their own water supplies and treatment plants.

What happened to doing your own business and letting professionals handle the rest?

Are you implying that the people working at university computing centers aren't professionals? I wouldn't. I would even count them as experts, since they usually are, or are in close collaboration with the people who do research on clustering techniques.

It's not like Sun expects us to pay through the nose for the processing power. They'd like to cover costs and have a decent margin, but probably nothing unreasonable.

$1 per processor and hour. The department I work at just bought a 440-processor grid for $286k. That means the whole cluster costs what Sun would charge for one month of the equivalent power.

Besides which, as another poster pointed out correctly, it's easier to get grants for funding equipment than it is for funding running costs.

The running costs of this cluster is zero for us. The computer center gets a share of the processing power instead, which they in turn either rent out our use in their own research.

Doesn't seem effecient

[adamruck]

Isn't the point of super computers, and clusters to do something really fast. This means having a custom system, and custom code, custom network setup, etc, for your problem.

If you can solve your problem in an hour anyway, I dont think its worth the time to have a grid computer do it. You might as well just run it on your own system, however big.

Re:Doesn't seem effecient

[Decaff]

If you can solve your problem in an hour anyway, I dont think its worth the time to have a grid computer do it. You might as well just run it on your own system, however big.

But if you can solve your problem in an hour on a grid, what is the point of having your own system for the rest of the time?

Re:Doesn't seem effecient

[Roadkills-R-Us]

This means having a custom system, and custom code, custom network setup, etc, for your problem.

Not true at all. Supercomputers have been used like the Sun Grid will be used for years. Theyve simply never been quite this cheap.

Even with custom software, you can develop it on a much smaller grid (two computers), develop your data set, then copy it all to Sun's grid and run it with the real data. Again, this has been done for decades with old style supercomputers.

I recall developing a FORTRAN program on my university's Cyber (early 1980s) on my personal account (we got a certain amount per quarter to do whatever we wanted with), then running it with the full data set on an IBM mainframe through a timesharing company for my customer. This paid for a quarter or two of schooling. 8^)

Re:Doesn't seem effecient

[adamruck]

Ok lets assume you have a small/medium size problem.

-rent a dedicated server for a month for 150 bucks, have control over your system, solve your problem as many times as you want, etc.

-pay a dollar/PERPROCESSOR/PERHOUR for a one time shot, while sharing resources with everyone else, while not having control over the system, etc.

Im thinking the first option would be really worth it.

The only situation where a service like this would be useful is if someone needs a HUGE amount of processing power, but only once every couple years. $1/per processor/hour adds up really fast compared to cost of hardware and power.

Re:Doesn't seem effecient

[ciroknight]

And you answered the reason for such a thing at the end of your post. Sure it's cheaper to rent a server for 150$ a month, but you're only going to get a recent pentium 4, maybe a xeon or a opteron. You could spend a little more and get a much bigger system, but this is around what you'll get, tops.

What if you wanted to do a research project on the fluid mechanics of the jetstream? (completely hypothetical, and just as an example of an operation that could be parallelized for speed). Your little 1-2 processor machine that you'll get for 150$ bucks is now going to take forever to do this, even for a relatively short sample. So, you now have an option of getting a 512 processor machine that can probably do the same amount of work in five to ten minutes, but since you're buying time in CPU hours, you can afford to do 6-12 times the sample space, for just 512 dollars, and can have it done in an hour, verses waiting a couple of months to do the same sample space, which would end up costing you relatively the same in monitary units, but a vast amount of difference in time. And as we know, Time == Money, so by blowing more time, you've actually spent more (especially if you're working for a company, which wants it done yesterday).

Just a thought/example.

Re:Doesn't seem effecient

[SunFan]

If you can solve your problem in an hour anyway...

You rent as many cpu-hours as you need. Need it by tomorrow, rent more CPUs; next week, fewer CPUs.

Also, this grid is running Solaris 10 on various Sun servers. I'm sure they can provide a cluster of Sun Fire 25Ks if you need to rent a few teraflops. When those teraflops aren't needed, their Grid Engine puts other tasks on those 25Ks, so they never go to waste.

Re:Doesn't seem effecient

[bitslinger_42]

Furthermore, many company applications don't have an easy way of scoping out future usage. I might have a job for a new product that's going to need 1000 CPUs for a week, because of the complex FEA involved, but once the product is developed and launched, I may only need a small fraction of that for ongoing improvements. I'm faced with the problem of buying 1000 CPUs, installing everything, configuring the OS and software, and hiring a staff of cluster-aware grunts to run it for its lifecycle, and then only using it for a week every two months. From an accounting standpoint, it is hugely expensive. It gets worse when you factor in issues such as keeping the hardware/software current and the fact that the problems tend to scale relative to the power available (i.e. engineers with no previous grid experience think up problems to be solved based on uniprocessor loads, but once they figure out how to parallelize well, the scope of their models grows). For a company whose business is not grid computing (i.e. most companies), maintaining a trained, well-adjusted staff to care for an expensive piece of equipment that's only used 26 weeks a year is not worth the money.

"Why build one?"

[dilvish_the_damned]

"Why build your own grid when you can use ours for a buck an hour?"
So I can charge 90 cents an hour.

Re:"Why build one?"

[lewiscr]

IBM posts as "dilvish_the_damned"?

Re:"Why build one?"

[SunFan]

If other companies enter this market, the prices will reach an equilibrium. Sun is setting the first bid at $1/cpu-hour. Also, if your $0.90 service sucks (you don't have a reliable power grid, good storage, etc.), it won't matter that you are cheaper.

Re:"Why build one?"

[T-Ranger]

Dont speek too fast. Many companies pay their electrical utility for an "interuptable" rate. Many of the chip fabs in silicon valley do just that, and get shut off first when there is a roaling blackouts.

I think it would reasonably well for a pay for CPU time type service. The customer has three choices
- build it yourself at >$1/hr
- pay Sun $1/hr, with no gaurentees of when it will finish as you arn't the only customer
- pay Sun's competitor <$1/hr, with no gaurentees of when it will finish as you arn't the only customer and the cluster may or may not be working today

If you are building your grid out of lesser components then Sun does, making it less reliable, you could also take that difference and buy more hardware. So you could do the work faster. So you might actually get the work done sooner, on average, with the cheaper solution.

Instructions:

[Anonymous Coward]

1- Insert coin here. $1 for each CPU per hour.
2- Insert DVD containing program in the tray on your right.
3- Please wait...

Old...saw this on a slashdot ad

[truz24]

I was reading about this, it was on the right and side where slashdot shows their ads... kinda funny

Talk about irony...

[naer_dinsul]

As I loaded Slashdot, there was an ad at the top for...
you guessed it...
The Sun Grid, available for $1/cpu-hr...

Are you sure Slashdot isn't selling advertising space disguised as news items?

Re:Talk about irony...

[zixel]

Ads???? What are those?

Adblock

Re:Talk about irony...

[Duncan3]

No Irony, this is how /. has been working for a while. It's smart.

Except for the constant dupes, most stories are now ads in one form or another.

Re:Talk about irony...

[Milton Waddams]

How is that ironic?

Re:Talk about irony...

[crunk]

*gasp*

I am shocked you would suggest such a thing!

Re:Talk about irony...

[k.ellsworth]

yup, adblock

How

[hey]

Does anybody know how you upload a job, get the results back, etc.

Odd Currency Exchange

[geomon]

"Why build your own grid when you can use ours for a buck an hour?" asks Sun's COO Jonathan Schwartz."

I feel a little wierd paying for my grid computing with venison.

It must just be me.

Re:Odd Currency Exchange

[PornMaster]

Will megabucks per second be a measure of aggregate CPU power that might be used to manipulate libraries of congress?

Re:Odd Currency Exchange

[geomon]

This reminds me of an old saw:

A billion years is unimaginable,
A billion stars is unfathomable,
But to Congress, a billion bucks is steak dinner for their contributors.

or was it:

A billion bucks here, a billion bucks there, pretty soon your talking about real meat!

Re:Odd Currency Exchange

[Roadkills-R-Us]

Whoa! Given the huge herds of deer we keep chasing out of our yard (massive nearby development having removed their quaint, forest homes), I could get quite a bit of free computing at a buck an hour!

How do you send medium sized animals via PayPal, though?

Re:Odd Currency Exchange

[wildwood]

Q: What's the difference between beer nuts and deer nuts?

A: Beer nuts are $1.79. Deer nuts are under a buck.

Thank you, I'm here all week.

Re:Odd Currency Exchange

[Paradise Pete]

: Beer nuts are $1.79. Deer nuts are under a buck.

It shames me to admit I liked that joke.

Like electricity

[wombatmobile]

Schwartz says computing will become a commodity like electricity.

Re:Like electricity

[valisk]

Not at that price it won't

$8766 for a single CPU year?

I could buy 17 Mac minis or commodity athlon XP boxes for that, and have enough change to pay for the electricity usage and a dominoes delivery or two.

$1 per CPU hour?

[Sensible Clod]

Thank goodness, I was starting to think they'd REALLY gone off the deep end...

Another Way

[SpottedKuh]

From TFA:

Mr Schwartz ran a demonstration of the service, showing how data could be processed in a protein folding experiment.

Of course, if your experiment is cool enough and academia-related, there are always other ways to get computing power. A similar chemistry experiment was performed using grid-computing in Canada, utilizing computing power from universities all across the country. http://www.cs.ualberta.ca/~ciss/

Now, granted this wouldn't be applicable to a lot of businesses, which is Sun's target audience. But the CISS project has a cooler name :)

I wonder how much to render a Pixar flick...

[PornMaster]

I'd be very interested in knowing how much it would be to render something like a Pixar all-CGI movie on their grid.

For all those who keep asking about cost-effectiveness... don't forget that when you rent from a utility grid, you don't have to worry about obsolescence - it's someone else's problems. You're not throwing out a bunch of P3s because P4s are available and better price/performance when the second project comes along. Renting CPU time is an operating expense. Running your own compute grid is both an operating and a capital expense.

Re:I wonder how much to render a Pixar flick...

[Effugas]

Rendering is only embarassingly parallel when done poorly. There's alot that doesn't change on a frame-by-frame basis, you know.

Re:I wonder how much to render a Pixar flick...

[alw53]

You also can expense the cost immediately rather than having to write it off over 7-10 years.
Unless you've gotten smarter (acquired intellectual capital) in which case you have to appraise and declare that, too!

Re:I wonder how much to render a Pixar flick...

[AFCArchvile]

Bit of trivia: Toy Story was rendered on 117 SparcStation 20 computers (87 dual, 30 quad). In that article they quote "800,000 computer hours" and "16 billion instructions per second", but remember that this is a press release from December 1995. Consider that in early 2001 Steve Jobs and a guy from Pixar demonstrated rendering Luxo Jr in REAL TIME!!!

Re:I wonder how much to render a Pixar flick...

[SunFan]

The market pricing for these things would take that into account. Customers who want to rent the resources but want to make sure it is a fair deal will inquire about the hardware (they could even track this in their account to double-check).

Re:I wonder how much to render a Pixar flick...

[prizog]

Toy Story, IIRC, was 10^16 flops.

Crandall claims that A Bug's Life was about ten times that.

Re:I wonder how much to render a Pixar flick...

[ms139us]

Mod parent up.

The real advantage to this is when you need the headroom. If 5% of the time your organization needs massive computing power, then you need to maintain that overhead the other 95% of the time when you don't need it.

With a utility solution you pay only for what you use, but have the headroom to spike when you need it. Very cool.

Re:I wonder how much to render a Pixar flick...

[tepples]

True, rendering 3D animation has some temporal locality, but if you have each machine rendering successive frames in a different scene, then there's a lot that does change from scene to scene.

Re:I wonder how much to render a Pixar flick...

[agurkan]

You missed the point. If you take correlations into account, for two successive frames you may not need to compute all pixels independently in the second frame. Deciding whether a pixel changes is a part of computation.

I'd imagine rendering each pixel independently in a given frame would be extremely impractical anyways, they are not independent and you need to distribute the information that can be reused for adjacent pixels to different machines if you treat them independent.

Re:I wonder how much to render a Pixar flick...

[danila]

According to Futures wiki article, most 3D films take about 3 hours per frame. The most complex special effects scenes in real movies (like the Superpunch in the Matrix Revolutions) take 20 hours.

An average movie is around 150000 frames. Assuming 3 hours per frame, this means that a full render will cost under half a million dollars. I don't know how many times you need to render the frame on average (taking into account previsualisation, test runs, reedits, etc.), but even if you need to do it 10 times (in full quality - you don't need a cluster to render a rough version), you costs would not exceed 5 millions. Not really that expensive. In reality, I think it would be closer to 1-2 millions. Which is peanuts for a Pixar feature film budget.

Re:Firefox is retarded

[Anonymous Coward]

It's so true. My first computer was a Commodore 64. Microsoft all the way.

Coincidence?

[dankinit]

The ad served by osdn for this article was this sun grid. Actually the ads appears all over slashdot and other sites now.

Anyways, here's the official sun page for the grid and a datasheet in pdf.

Because its Sun...

[suso]

...there will probably be several research businesses and institutions that will be tricked into thinking that they need to use this pay by the hour grid instead of making their own which would be more cost effective.

Re:Because its Sun...

[CommieOverlord]

Then those businesses/institutions didn't their research. No, this isn't effective (for cost or other reasons) for all cases, but there are some cases where it will be. You can't blame Sun if customers failed to go with the solution that best suited their needs.

Re:Because its Sun...

[AviLazar]

I doubt there are many upper level CTO's (this is a CTO decision) who are not knowledgable enough to realize that this service is only useful if you are not planning to create your own system because you do not need such a long-term device.

We are not talking mom-n-pop shop here...we are talking higher end. So a university who decides to use this is probably doing so becuase it is just one or two short-term projects...but if that university plans on doing many future computing projects (not unlikely) they will probably build their own system (using the college grad students as cheap/free labor)... People on the upper end of the decision making process are not that dumb (though exceptions abound).

Re:Because its Sun...

[SunFan]

Imagine the electricity, support costs, and staff for maintaining your own cluster. If you aren't keeping it running tasks 24/7/365, renting the time from Sun could be very cost effective. They are basically renting out their farm of thousands of CPUs as a compute hotel...by the hour.

Re:Because its Sun...

[LurkerXXX]

Only if they are going to make good use of your own system. For others Sun's deal will be very cost effective.

Suppose you do some fancy quarterly statistics/forcasting analysis that take 20 hours on that 1,000 CPU cluster( which would have taken you 833 days to run on a single CPU machine). That might be your only major need for intense CPU power.

That's when you want to use these clusters. For the price of $20,000 each quarter you can avoid the cost of a 1000 cpu cluster (which will be several hundred thousand dollars at least), plus building space, maintenance, cooling, power, administration cost, etc, etc, etc. Plus SUN will likely be upgrading their clusters regularly, and that would be an additional cost to you to keep upgrading your own cluster. Sun's deal makes a lot of sense for occasional use high intensity jobs.

If you have enough researchers doing enough things to keep one busy most of the time, then yes, you are right, it would be cost effective to build your own. But there are going to be a lot of places that don't have such a high continual need.

Re:Because its Sun...

[nelsonal]

I didn't realize the red light district required such computing power, what do I know.

Ya Think??!

[handmedowns]

Considering these ads have only been on slashdot all day long..

I love how a good deal of the slashdot community are supporters of open software and standards.. and yet every day that I read a story on slashdot, about how so and so country / state / organization is implementing foss, right SMACK in the middle of the page is an add from MS on the "Facts" about TCO for Windows vs Linux.. or "Facts" about Performance..

Should be "News for Nerds, Irony that matters.."

Re:Ya Think??!

[cgori]

Those of us "implementing foss" use Firefox with Adblock and never see what you are talking about.

Duh.

It's too expensive.

[sbaker]

I can buy a pretty decent computer with a modern CPU and lots of RAM for under $500 - maybe as little as $300.

So unless I need my results very soon after posing the problem, I'm better off spending $500 bucks on a PC and running my problem for 20 days than I am buying 500 CPU-hours from Sun and getting the answer back very quickly.

But Sun must have to schedule their system - and you have to go through the grief of sending your program to them, getting it to work on their grid, paying for it, etc, etc. So you know it's not going to be available on-demand, *instantly* - so you might have to wait several hours before they can schedule your task. This facility is only going to be useful for things that would take an eternity to run on a single PC.

Even if I need the results quickly, unless this is a one-time problem, I'd be better off buying a pile of cheap PC's than using Sun's facility. If I need to run a 100 CPU/hour problem often, I can justify buying a $10,000 20 PC cluster for just 100 runs.

Bun if Sun's niche is big problems whose results are needed quickly *AND* which are not run frequently - then there is still a problem because you just know it's going to be quite a bit of grief to get your code ported over to Solaris (or whatever they are running) - to get your data onto their disk drives - to get the results back. If you only run this program once - then that overhead will kill you - and you'd *still* be better off buying your own systems.

FWIW: IBM offer a very similar service - with very similar problems over pricing.

Re:It's too expensive.

[Necroman]

I've done some work in the HPC (High Performance Computing) field, and for a lot of the applications in research, you only need that much CPU power once. I did some coding that I had run on NERSC, and I can see the use of this for private companies (NERSC is owned by the Department of Energy).

This is what I've seen of people doing research type work. The researchers will have a smaller cluster always available to then (2 to 16 node setup), that they use for all their initial development of the application. But when they need to run the program for real, doing their full calculation, they would farm it out to some big system like NERSC. The scheduling systems these kinda of systems have tend to do a really good job at scheduling workloads, and the wait tends to be minimal.

I think this is a good idea for companies that don't want to build their own grid. The cost of the computers might not be a lot, but if you have an application that requires a lot of communication between systems, you need a really good interconnect system (such as InfiniBand) cost a lot of money to setup. You could spend as much on a good interconnect system as you do on all your computers, if not more.

Re:It's too expensive.

[CommieOverlord]

Are you sure?

You haven't factored in fast networking (IB, Myrinet, good GigE, switches), maintainence, space, electricity, A/C, security, storage, backup, manhours. Clusters generally use Xeon/Opteron/Itanium, wondered why?

Re:It's too expensive.

[nodrogluap]

You have to look at the total cost of ownership though. Suppose you are a business manager in a firm, in addition to the cost of the buying machines you have to consider:

-Man hours spent unpackingly and assembling the cluster (hartdware and software installs)
-Networking infrastructure (the switches will cost you a few thousand at least for a decent cluster)
-Physical space to put the machines (no one wants this space heater next to their desk)
-Electricity cost of running the cluster and cooling the space
-Man hours spent getting patches, checking for hardware issues, etc. (i.e. sysadmin time)
-Man hours spent adding this equipment to inventory lists for the accounting department.

Plus other stuff I haven't thought of yet. These costs (especialy people time and space logistics) are often a bigger factor than the commodity equipment cost. Plus, if your cluster sucks, you have no one to blame but yourself! If your core business is selling widgets, giving a company $1 per CPU hour to process a big important job doesn't sound like such a bad idea, unless you're running it over and over again...

Re:It's too expensive.

[Necroman]

Here's the fixed links:

http://www.nersc.gov/

http://www.intel.com/technology/infiniband/

Re:It's too expensive.

[SunFan]

"It's too expensive."

You aren't their target customer.

Re:It's too expensive.

[drew]

So unless I need my results very soon after posing the problem, I'm better off spending $500 bucks on a PC and running my problem for 20 days than I am buying 500 CPU-hours from Sun and getting the answer back very quickly.

or maybe you are doing a complex mathematical problem that needs to store lots of intermediate steps in memory. even if you don't care about how long it takes to run, there are some applications that will never run on cheap p3's even if you give them until the end of time.

that's the market sun is after with this.

Re:It's too expensive.

[bitslinger_42]

And its not just the CPU speed. The memory's got to be fast, the backplane needs to be able to move huge chunks of data between memory, CPU, and storage, the disk needs to be fast enough that it doesn't take days just to read the initial dataset. Stability is important, too. Imagine how you'd feel if you'd spent millions of years running a complex program looking for the answer to Life, the Universe, and Everything only to find that the computer was blown up at the last second? The supercomputer farms have the benefits of economies of scale working for them. Since they're dealing with a variety of customer's needs, the equipment is always in use. The support staff has enough different things going on that they're kept entertained (ever tried dealing with a pissed-off, bored support guy?). All of this combines to make renting by the CPU hour a very attractive premise to anyone who isn't in the business of doing pure research.

Re:It's too expensive.

[CEODRI]

Actually the networking kills you in the cost, unless you have an embarsingly parallel problem. Otherwise dicestion of bandwidth kills you. Infiniband is very expensive.

Re:It's too expensive.

[danila]

Forget IBM, I wonder how Google will price their version of this service...

If you can't sell it, rent it

[gUmbi]

So Sun's finally found a use for all of their spare inventory.

It's funny how old ideas become new again though...Is Jonathon Schwartz/Sun trying to become the new Ross Perot/EDS?

Re:If you can't sell it, rent it

[Animats]

So Sun's finally found a use for all of their spare inventory.

That sounds about right. Scientific time sharing hasn't been a good business model since 1980. If you need heavy compute power, you get your own cluster. If there was a viable business model in this space, hosting companies would be selling this as a service. They already have the right infrastructure.

For a while, it looked like commercial render farms might be a viable business. But today's stats at ResPower read "Running frames: 3, Waiting frames 0", so only 3 of their 500+ computers are active right now.

The "use spare cycles on other people's PCs" model works fine, if you're a spammer or an adware/spyware company. But nobody seems to be paying out money to home users for spare cycles.

Re:If you can't sell it, rent it

[SunFan]

If there was a viable business model in this space, hosting companies would be selling this as a service. They already have the right infrastructure.

Not quite, as they aren't running an N1 Grid like Sun is. From what their web site says, I estimate they are running a Grid Engine allocating out to Containers on their servers. That means you can rent as much or as little of their servers as you need, and receive complete isolation from other people's tasks.

Re:If you can't sell it, rent it

[dkf]

If there was a viable business model in this space, hosting companies would be selling this as a service. They already have the right infrastructure.

The problem is the software vendors. They usually (well, at least at the high-end; we're not talking consumer stuff here) want to know the names of all their users (not just customers, but the names of the people within the customers' organizations who will be using it). This, combined with the fact that applications have to be tuned carefully to take advantage of the hardware available, means that it is very difficult for hosting companies (the term you are looking for is Application Service Provider, BTW) to get all the bits and pieces together to make a viable large-scale business case.

In practice, there are a lot of other problems too (e.g. customer-to-ASP networking is a known serious issue for serious stuff) and there's room for a lot of work building this sort of thing into something truly viable.

Re:If you can't sell it, rent it

[Animats]

Actually, no, it's not a software licensing problem. Look at Respower, the rentable render farm, again. They're licensed for all the major rendering engines: Maya, Mental Ray, Lightwave, 3DS Max, etc. (Not RenderMan, though.)

This is in an industry where the software works well on clusters and customers are used to outsourcing. Yet it's a business disaster: ResPower's load today: 1 frame in progress, 0 waiting, 500 machines available. 99.8% of their machines are idle today. Worse, if you see only one frame in progress, it's probably their freebie demo. Anybody can create an account and render one frame on one machine for free.

ResPower used to be busier. Maybe it's a slow week. Or month. Or year.

Despite the "grid enthusiasts", grid computing, as a business, is a disaster. You can build it, but they won't come.

Where's the News?

[ohuf]

in fact, they announced this model almost half a year ago... And please note that this includes no grid soft by now.

Sun offers energy by the hour

[GillBates0]

Sun offers energy by the hour

The Sun has launched a pay-as-you-go service which will allow customers requiring huge solar power to rent it by the hour.

Solar power costs users $1 (53p) for an hour's worth of light and heating power on land covered by Sun.

So-called fusion reaction is the latest buzz phrase in a solar system which believes that solar energy is as important a commodity as hardware and software.

The Sun likened fusion reactions to the development of electricity.

'Buck an hour'

The system could mature in the same way utilities such as electricity and water have developed, said the Sun.

"Why generate your own power when you can use ours for a buck an hour?" he asked in an address launching Sun's quarterly solar eclipse event in the center of the Solar System.

The star will have to persuade the entire galaxy to adopt a new model but it said it already had interest from planets in the milkyway, andromeda and B53 stellar clusters.

Some of them want to book capacity of more than 5,000 TeraWatts each, Sun said.

Mr Sun ran a demonstration of the service, showing how fusion could be performed on elements.

Hundreds of atoms were fused simultaneously, generating energy for a few seconds each.

Sigh....too much time and an agile mind.

Karma Grid API....

[conna01]

I'll donate my domain's idle CPU tasks to somebody elses grid in exchange for whatever task i might need later. Somebody design the

Re:Karma Grid API....

[conna01]

The book keeping might be the first task the grid is good for.... Also why stop at CPU power? Share disk space. Have an internet based raid setup.

Re:Karma Grid API....

[AaronGTurner]

I think we will see more of 'trading grids' in the future. The White Rose Grid (Yorkshire, UK) is a form of such a grid in an academic environment.

In addition there are the usual array of volunteer CPU cycle scavenging Grids for good purposes : protein folding @ home, set @home, IBM's Community Grid and so on.

What is THE CPU ?

[nomad63]

by means of $1 per CPU hour, are they referring to a p4 or a sparc processor ? At what speeds ? This can make a huge difference how quickly your CPU intensive problem gets solved.

Re:Ad before stories?

[PinkX]

Man, I just clicked on the details of this story to post *exactly* what you said ...

too late for RFID cracking proof

[essreenim]

Remember the guys who used 32 FPGA boards (about $3200) in parallel to crack 40bit RFID keys. It must have taken forever to program the logic for those. Just think, if they were able to use a grid instead?

Still, I do wonder how much freedom people will have to carry out private experiments. Not much probably...

This is so retro!

[jhobbs]

The Seventies came back in fashion, why not in computing.

Re:This is so retro!

[ajlitt]

Everything Old Is New Again. Next thing you know, Sun will try pushing thin clients as the next big thing. Again.

Re:This is so retro!

[mbrewthx]

How true!!! The more things change the more they stay the same.

Oh I miss managing the old PDP-11...

Okay so now the Computer is the network..
But a few years ago Sun told me that I didn't need another box I needed a network.
Ahh I'll stick with my linux box

Re:This is so retro!

[davecb]

It reminds me of Plan 9, where I can say "cpu /big/complicated/job &" and shove it off to a compute server.

Mind you, Plan 9 From Outer Space is seriously retro!

--dave

What's with all the stock pictures?

[gotr00t]

Does anyone notice how on the BBC article, both the pictures on the page have almost nothing to do with the article itself? The picture of a guy in a lab coat taking his glasses off is kind of a stretch, and the other picture, some old IC's, is totally irrelevant, considering the caption is about how sun offers tons of computing power, and virtually no modern microprocessors have the DIP package anymore.

I think that BBC should stop using stock images if they don't actually have images that pretain to the story. I mean, this isn't some high school jornalism class.

Re:What's with all the stock pictures?

[__aazofn1209]

Well, as someone else pointed out, they don't specify what kind of CPUs you are renting. I think it's clear from the pictures that you're actually renting some 7400 series TTL chips.

So good luck porting your bioinformatics app to the 74LS00 Quad 2-Input NAND architecture.

But I, for one, welcome our new 7404 hex inverter overlords. My collectors are open.

Okay, I'll stop.

Re:What's with all the stock pictures?

[iggymanz]

74LS00 ??!!!

I expect to sink full TTL loads on my $1 / hour / chip grid, none of this low power schottky crap, thank you very much.

Capitalism

[No Such Agency]

Hey, if they have CPU's and want venison, and you have venison and need CPU's, isn't that how capitalism is supposed to work?

Not a bargain compared to colocation prices

[Bruce Perens]

At $744/month (24 * 31) you can buy at least three of the best Intel or AMD CPUs in a dedicated server along with the rest of the system including disk storage. I used to render movies. They were so I/O intensive that it did not work to have the data - including a lot of photographic textures - all live "on the network". It had to be on the local disk.

Bruce

Re:Not a bargain compared to colocation prices

[Bruce Perens]

There is a maxim in film computer graphics: A frame always takes 5 minutes to render. If you add faster networking, better CPU, more memory, etc., the animators use it. They add enough complexity to the scene that you are back where you used to be.

So, instead of anything working faster we get realistic looking hair, which we were formerly unable to do.

So I'm not sure where we stand today regarding local disk.

Bruce

Re:Not a bargain compared to colocation prices

[Bruce Perens]

I have a Dell dedicated server that I've never seen, in a rack somewhere in Texas, with really great networking and professional management. It costs $250/month. The folks who did my server recently did a 500-node install for a grid customer. It's not a partition in a Sun N1 grid. It's a high-end Intel CPU, a lot of memory, and reasonably good I/O, and I don't share that with anyone. I'm not clear what the bottom line difference would be between my system and the grid.

Bruce

Re:Not a bargain compared to colocation prices

[SunFan]

I'm not clear what the bottom line difference would be between my system and the grid.

Sun could double or triple or whatever your CPU/storage allocation in a moments notice and your bill scales linearly. Their system is completely virtualized.

Re:Not a bargain compared to colocation prices

[Bruce Perens]

That's a good point. If I had to crunch to render a film in time for Christmas release, I might go for 10,000 nodes on a moment's notice. Indeed, Sun once did lend Pixar a whole bunch of computers on very short notice, to do just that.

So, this could be a good proposition for bursty use. If you can forecast your needs, you might at least provide for your baseline capacity in a more cost-effective manner. But you could supplement that baseline with grid services.

Thanks

Bruce

Ammendment

[Uber Banker]

Read "under the reigns of Linux" as "under the reigns of IBM". Here follows the ammended post:

Haha, how cynical! But given they're in competition and not collusion with IBM (for now) plus the 'traditional solutions', perhaps that is second order in their minds.

Where this is going of course, as the article touches on, is the commoditisation of raw computing power, making it a product like iron ore, coal or oil. Genericism will, IMHO, be a really interesting force behind evolution of computational techniques over the next 10 years.

With genericism perhaps there can be no monopoly in provision of computing power, or more specifically for Sun, there is genericism until we get to the buts-and-bolts Solaris and then we get a monopoly in proven scalable OSes and as long as they can hold off Linux snapping at their heels under the reigns of IBM, then they'll have their monopoly and that is step 4; which leaves holding off Linix presumably as step 3, just you forgot to denote it '???'.

Buzzword++

[Duncan3]

1. 50 years of rented computing history
2. Add a Buzzword and slap "NEW" on it.
3. Press release bonaza!
4. Profit?

Yeahh, marketing

[essreenim]

Microsoft can rent the grid to predict the most popular desktop Linux distro in 5 years time, and then focus on incorporating it's present features into Windows and patenting any useful features for the distro in the future.

OR

they could use it to predict how dummed down people will be in 5 years and use that to push the maximum user-friendliness into Windows..

Ohhh, I'm just joking really....

Cool...

[Anita Coney]

Can you imagine renting a Beowulf cluster of those babies?! That'd be awesome!

An Internet beowulf

[wikinerd]

Now we need a Linus Torvalds version 2.0 to build a free open-source P2P Beowulf cluster over the Internet.

Imagine playing Doom 3 on a P-II, with the graphics being rendered by an Athlon64 somewhere in the Internet.

Now that many computers are connected to the Internet with fast DSL connections, it would be very beneficial for all if someone could start such a project.

The basic software already exists and it is in the public domain: MPI.

I explain this idea in more detail on my blog.

Re:An Internet beowulf

[Anonymous Coward]

Imagine playing Doom 3 on a P-II, with the graphics being rendered by an Athlon64 somewhere in the Internet.

Alright... *closes eyes* I'm seeing a world without bandwidth restraints.

Please; Look into the grid and clustering research before you make wild claims.

Re:An Internet beowulf

[Anonymous Coward]

Further, you completely ignore security. I could send porn back instead of Doom graphics.

Please; Just sit down.

Re:An Internet beowulf

[conna01]

if it was doom3 anything i could see on my screen would be an improvement.

Re:An Internet beowulf

[Hydrogenoid]

Grid computing tends to have relatively high latency between nodes, usually, (and worldwide ones are obviously even worse in that respect) and quite small bandwidth, which limits its usefullness to a relatively small category of problems. (think SETI@Home kind of problems: not a lot of data to transmit back and forth, loads of cpu power required)
And interactive real-time rendering in a game definitely isn't one of those, to even get a compressed video at 720x480 resolution, compressed, (DVD Video) you need a 9.8Mbit pipe...
Just try to use a remote desktop application or export an X session, and you'll see what kind of speed you get...

Re:An Internet beowulf

[SunFan]

Do you think a company with a dataset full of trade secrets would run their job on your Internet P2P Beowulf cluster?

Re:An Internet beowulf

[mollymoo]

Now we need a Linus Torvalds version 2.0 to build a free open-source P2P Beowulf cluster over the Internet.

Imagine playing Doom 3 on a P-II, with the graphics being rendered by an Athlon64 somewhere in the Internet.

I bet a global P2P grid could get millions of frames per second in Doom 3! You'd be fragged before the first frame was back, but who cares? Once you'd spent a fortnight downloading all those 1600x1200 bitmaps over your 33.6K modem you'd be able to see how you died in incredible resolution!

Time-sharing is new(s)?

[still cynical]

How can we submit our programs? Punchcards are OK I guess, but do they have magnetic tape?

Helping the little guy?

[PxM]

It's nice that there might be a unidirectional market for this, but what about a shared system where you trade your idle time for the right to use someone else's computer during your peak time? End users tend to want a lot of peak power but little average power. Grid computing would let you spread the peak out like a capacitor. If the average user needs N times their CPU to run an application at realtime speeds and they only use this peak for 1/N parts of the day and are idle the rest of the time, you can have N users from around the world distribute their load onto each other's machine.

This would work for applications that don't have low latency (e.g. FPS games would be out) and applications that don't have high bandwith (e.g. compressing movie files) but it would work for things like semi-professional graphic artists who can't afford renderfarms. By letting other people use their machine for 1-1/N of the day, they get a near N increase in computing power at peak times.

You could have every machine in a company run a VM at the lowest priority and share the cycles for this VM with other companies around the world. You would get some use out of all that idle time on modern machines. You could also have a company's engineering department run their simulation on the company's accounting department's machines since those machines are probably idle all the time and the simulation machines have a high peak CPU requirement.

This would even work well locally if IBM's Cell technology delivers. Your TiVo can offload the processing for digital compression into the chips in your PC, DVD, and other high end processing systems as needed or vice-versa the next time your PC needs to compress a bunch of files.

Given all the idle time on home machines, this might be a bigger market than supercomputer grid clusters.

--
Free iPod? Try a free Mac Mini
Wired article as proof

Re:Helping the little guy?

[AaronGTurner]

This is certainly being done in academia already although generally without the strict QoS and other legal issues that would be required between commercial entities.

With regard to the scavenging there are a number of tools that can be used such as ones from Entropia, Condor, Mosix, Inferno and Sun Grid Engine to name a few, plus some stuff coming out of IBM's Alphaworks. With regard to virtual machines there is also work going on with regard to this, Solaris 10 virtual containers and VM Ware being two examples, and also the ability to run Linuxes within Linux. At Duke University there is a project to allow some form of automatic creation of such virtual servers to allow on demand provisioning of specific environments. I also wrote a paper for a Sun HPC conference covering aspects of this.

There is a lot of stuff happening in this area. We're not there quite yet, but I would expect to see some robust solutions to this emerge over the next three years. There will still be issues to solve in terms of the subtleties of balancing policies and making the best possible scheduling decisions plus the rather complex issue of legal and QoS agreements.

Prof said..

[Misch]

Last night, my elder Computer Architecture professor was talking about how come computers he worked on in the 1970's could go for well over $5000 an hour.

"Today, by comparison, it's essentially free" he said.

Hmmm.

[eclectist]

This reminds me of a friend who had a weekend solo-IT job that sucked rocks. His only solace was that he had a number of knoppix disks and would cluster all the unused boxes in his cube farm and use the processing power to render graphics. He thought about hiring it out, not by hour but by project, but we both agreed that no one would want the service...

I wonder if he's seen this article

Re:A possible use in combatting spam

[PinkX]

Except that almost all of the computers used to send spam are compromised zombie hosts all over the Internet, varying from XP machines to poorly secured (or not secured at all) linux boxes, whose owners have no idea at all that their computers resources and internet connections are being wasted in sending thousands of messages an hour. Regards,

Re:My question - what kind of processor

[rindeee]

Okay, so who says it's not? Sun sells many an Opteron based server right along side their Sparc brotheren.

License issues

[josecuervo]

How will this work with license issues. Companies can't just ship licenses for pieces of software, such as vlsi synthesis to an offsite grid. The software makers generally aren't ok with things like that.

Re:License issues

[AaronGTurner]

Increasingly it seems that companies are understanding Grid. I've certainly bumped into some companies selling software used in computationally-intensive applications at Grid conferences.

One dollar per hour, Jesus.

[alw53]

I can remember blowing $200 per minute on the Univac 1108 at Georgia Tech, when my program got into an infinite loop.

Its not just what the boxes cost

[tallsails]

Our datacenter costs include UPS cost, cooling costs, rack space, operators, 24x7 on call techs, backups, off site backup storage, which multiple the CPU box cost by several factors.

The current big users of this are financial institutions running monte carlo analysis of stock and commodities markets - they use thousands of CPUs at a time, every day.

The other nice feature to this, is you get glass-house data center capability, but can turn it off, which is not as easy with your own lease agreements, etc...

To paraphrase Atari: Do the math.

[Rhys]

Do the math. (anyone else remember those old Jaguar ads?) A basic Xserve runs right at 3 grand. Upgrade that some (memory to the dual proc 2.3 ghz model), throw in overhead cost for a fast network, some storage, head nodes, etc... and how much are you actually talking in terms of (amortized cost) for an xserve in a cluster? Let's say 5k (an overestimate I'd say, but I didn't go try to price out a fast interconnect unlike hitting apple's page for an Xserve price real quick).

That's for a dual processor box too, so cost per processor is more like 2.5k. So, ask now how long you have to use that machine before you've made the cost of using sun's service?

Let's simplify: 4.8k for the xserve, 2.4k per processor. That's only 100 days running flat out. A third of a year (less!). Yeah, you aren't likely to have the machine loaded 100%, 24/7 yourself, but I'm sure when the Xserve cluster here at UIUC is officially open it will be pretty much running 100% 24/7, because the old PC cluster it's replacing was always running nearly flat-out. The apple replacments are currently running about 40% load and they aren't even open to "the public" as it were. (with 10% additional being used to tune the cluster by our resident parallel expert)

Yes, there's system administration overhead, and cooling and power and all sorts of other things you need to take care of too. And, for places that aren't a university that don't have the resources to deal with that, Sun's grid may be a good deal. Or places that won't load the machine up 100% 24/7.

But for a large compute cluster that you're going to keep pretty busy, you can do better.

Yet another clueless marketing abuse of Grid

[deadline]

Having a pile of servers does not mean you have a grid. To use the electrical analogy, there are different electrical grids that are not compatible i.e. European and US electrical grids. If you plug into the wrong electrical gird your device does not work. The same is true for computer grids. they are designed to service specific applications. Getting all the details right is hard -- that is what the Globus project is all about.

Now what IBM has been doing is not Grid. You basically rent a machine for a certain amount of time. You actually start with a small test cluster, then when it works, your "image" is then transfered to the real thing.

Grids are designed so that everything you need for your code is on the grid you are using (including data). On-demand is renting cycles.

Nice try, but this market's established already.

[Mindwarp]

Sun are entering this marketplace late in the game. Not only that, but the pricing of $1 per CPU hour isn't even close to competitive in the current market. There are plenty of vendors that offer grid capacity on a 'peak demand' basis for a lot less than $1 per CPU hour.

fees

[SecretSqrl]

Do they charge you a fee if you return it late?

Mhz Currency

[ElitistWhiner]

It's not surprising that computing has returned full circle to timeshare... its that cycles have achieved currency.

Sun are trying to assign a "store of value" for the CPU cycle at $1/CPU/Hr. For Sun to achieve the promise Mhz Money, ie. store of currency, their clusters must produce.

Despite the fanfare, marketing and posturing, the only thing Sun are storing are cycles. This is a pure "warehousing" scheme, storing cycles and charging rent for its use.

Interesting if only they leverage Sun into the branded warehousing business...

Re:click

[CharlieHedlin]

Depending on how they bill this could be very usefull for chip simulation as well. I setup a 12 node cluster at a previous job and we had jobs that took 10-15 hours. Although I have to admit that a major factor in the size of the cluster were the simulation software licenses.

Doubt it is completely sustainable...

[Junta]

http://www.eweek.com/article2/0,1759,1758758,00.as p
is the article I read

I don't think 'betting the farm' on on-demand computing is a good idea for sun. The market is there and has money, but I don't think a company like Sun can sustain itself on almost exclusively this.

IBM does this with their servers, but if they didn't have a bunch of customers also buying the systems whole working to help subsidize research and development of the servers, not to mention manufacturing the systems in volume, the revenue from on-demand would not be sustainable.

I found the comment about IBM pensions amusing, but the reality is that while IBM does have armies of people getting paid keeping 1$/cpu hour from sustaining the company, the statement ignores how much broader IBM's scope is compared to Sun's present and plan.

Unanswered questions

[ZorbaTHut]

A lot of people assume "CPU power is everything!" but that's not really true. Questions that haven't been answered, that are absolutely vital to how good a deal this is:

How much RAM per process?
How much connectivity between systems? (Gigabit switch per rack? Better?)
How many available disk space, and how fast can it be accessed?
And, of course, how *fast* are the CPUs?

I'd want a LOT more info before I even considered using that service, and I just can't find where the info is listed.

How fast is each cpu?

[jrushton]

Ok. Following on from that, assuming you were operating a terminal...

If you used it 12 hours a day of straight 100% CPU use, for 365 days a year itd cost:

$4380

If theyre upgrading, and maintaining...
You would also spend an awful lot less that that because you wouldnt be running the cpu flat out.

Do thet sell disk space too?

Re:How fast is each cpu?

[Handpaper]

Do they sell disk space too?

Yes. $1 per GB, available, backed up data-centre style, warrantied etc.

I read TFA yesterday, and did consider submitting it, but I don't like repeating myself.

So, it seems there is an emerging rental market in CPU time. Maybe I can work out a way to make my E4K pay its way?

Wait and see

[tepples]

Companies can't just ship licenses for pieces of software, such as vlsi synthesis to an offsite grid. The software makers generally aren't ok with things like that.

Yet. Watch publishers of proprietary programs start offering "Grid Editions" of their software, licensed specifically for upload to a rented cluster. Can you imagine a Beowulf cluster of revenue streams?

Re:Portability - Mobility

[99BottlesOfBeerInMyF]

but it can be "test driven" for literally any budget, to the dollar... can pay a minimal fee (like a dollar) to install its "spec test" code on each available grid

I'll believe that when I see it. I imagine everyone else here who has dealt with Sun before is thinking the same thing. "Hello Sun, anybody there, I want to spend 5K on a workstation. Will anybody sell it to me? Anyone? Call me back? Are you guys awake?"

Maybe they have gotten better over the years, but for those of us who had trouble making a small purchase (5k) in the past it is hard to believe their sales people will call you back to set up a $10 test. I think you are dreaming.

Re:Computing power as a commodity?

[Derkec]

That's exactly what Sun is trying to have happen. They don't neccassarily want to be the utility long term, but they would like to supply the technology to the utilities.

This is yet another extension of the old standby slogan "The Network is the Computer." You push your data out onto the network (utility grid) and get your results back.

Sliding pricing scale?

[grahamsz]

Maybe pixar could offer to pay a lower rate for a lower priority of job. That way they could suck up all the spare cycles whenever others weren't paying.

Not clip-art

[andy@petdance.com]

check-out-the-clip-art-scientist dept I think you mean "stock photo."

Security

[Migraineman]

I'm running a fluid-dynamics simulation for a super-secret aircraft/submarine/time-machine. Can I guarantee the security environment for my processing? No? I guess I'll be building my own cluster, thankyouverymuch.

Re:Portability - Mobility

[SunFan]

I've made small purchases from Sun a few times (less than $100, even). Use their website. It's just like any other mail order business like CDW.

Re:Portability - Mobility

[99BottlesOfBeerInMyF]

It has been years since I purchased anything from Sun, but at the time you could not purchase directly from the website. It would not even tell you the price of a workstation, only connect you to a sales agent who would haggle with you when he got around to it. Given the fact that this service will probably require personal interaction, I fear it will be just as bad. Maybe it will not. Maybe they will even have it all nicely automated. We shall see.

Re:grid computing

[Mindwarp]

Grid computing always seemed like a stupid idea to me.

If by stupid you mean 'cost effective', 'scalable' and 'provide a high return-on-investment' then yes, you're completely correct.

Computing resources are cheap; it's managing them and setting them up that is expensive. Dedicated machines are cheaper to manage and easier to set up than grids of under-utilized computers.

You're right about the up-front cost of the hardware being the smallest expenditure over the lifetime of the machines. The biggest cost by far is typically the cost of actually housing the machines in an air-conditioned data center. In some situations it's nice to be able to push that long-term cost onto someone else.

You also seem to be a little confused about grids. It's not a requirement that servers/compute units be permanent members of a processor grid. It's entirely possible to have regular servers that are pulled into an 'amorphous grid' during periods where their primary functionality is in low demand. In fact, from what I've seen in business this is typically the way it's done.

The ultimate goal in many grid computing situations is to maintain as close to 100% CPU utilisation as is possible. By making one departments unused CPU capacity available for general use, business centers can get closer to this ideal (lowering future hardware purchase requirements and making the fixed datacenter housing costs more palateable.) Rental systems like SUN's recently announced one are typically used to allow customers to cover short-term peak grid demand without the need to invest in new hardware that, seeing that it would only be required during those peak hours, would otherwise lower the overall CPU utilisation statistics.

Either that or you're in the wrong thread...

[Roadkills-R-Us]

Not to be a smart aleck, but so what?

Do you really think the work you do has anything to do with the work everyone else does? You might as well claim there's no need for big trucks or railroads, because you can buy an economy car, put a *really* low gear in it, and eventually you can deliver anything, anywhere.

The fact that you only need to haul a small trailer now and then has nothing to do with the fact that someone else may need to haul enough steel to build a World Trade Center. And if you think you're going to haul that steel cross country with a state of the art Kia, you've lost your mind.

We have 200+ systems in our sim farm at an average speed of 1200MHz. We really need more power, but can't afford to replace all the systems every 1, 2, or even 3 years. And if we had to run everything on one system, even the fastest PC available today, we'd be out of business in short
order because our schedule would be slipping faster than we could update the customer.

If this Grid of Sun's ran Linux instead of Solaris, I'd be seriously looking at it. As it is, we have been switching from Solaris to Linux because licensing costs for the tools we use are much lower for Linux.

Re:Either that or you're in the wrong thread...

[cgori]

And this is the true problem. I also run a sim farm (but not as big as the parent's). State of the art linux machines can run 2x to 3x faster than a top-of-the-line Solaris machine, and the licenses are cheaper in some/many cases.

If Sun starts to offer Opteron grids, give me a call.

This isn't a replacement for colocation

[ievans]

Colocation and renting time on a grid are not the same market. Consider:
"Renting office space is not a bargain compared to buying real estate." Well, in some cases that's true, but it's by no means always so. The advantages of renting office space are quite similar to renting grid time.

So how does it work again?

[mnmn]

I just visited the site and tried on a few of the PDFs. It seems to be aimed at higher markets, and maybe they wont rent out single processors. I was hoping I could use this as a webserver on a single CPU if cheap, but $700 a month isnt.

I was compiling kernels for an embedded project a while ago, and many flags of the linux kernel makes it crash, especially with the -tiny patches. I have a PIII at home, so recompiling after each crash was a royal pain.

So the idea was I'll rent a grid of CPUs, and iterate through all possible flag combinations (gcc flags, config options, patches even) and download a large number of small compiled kernels, and try the successful ones out for size, to find the smallest kernel I can compile for the project. Unfortunately it gets way too expensive, and I'd much rather replace my motherboard and cpu with an Athlon64, or two, and leave the thing running for a week. Or a year.

The grid isnt for everyone. Its for people who know exactly what they want, know about the parallelism of their problems and how to make highly threaded apps, or distribute the execution across CPUs. Its not for you and I.

Re:My question - what kind of processor

[AaronGTurner]

" They're simply the worst for scientific computing"

If you are talking about Sparc architectures a lot depends on how you optimise your code. You can take code that runs moderately well on a high-clocked x86 architecture and have it run like a dog on a Sparc. However if you have floating point mathematics and tune the code to dispatch work to each of the four floating point units and keep those pipelines full you can get massive speed ups. Lots of the techniques will also provide lesser speed ups on x86 architectures too, but the Sparc has the advantage of a lot of floating point horsepower, but you do have to do a bit of extra work to get the best out of it.

As part of the Forte compiler set (I can't remember the latest Sun ONE name for this) you get a combination of text-based and graphical tools (depending on what software set you have) that allow you to analyse the code performance in depth and tune it. Sun also runs some excellent training courses in this area.

Architecture

[vlad_petric]

I agree that, as long as you can *vectorize* your scientific code, SPARC should work well. Otherwise, keeping the fp units busy can be extremely tough, even if you work very hard . Why ? To put it simply, SPARCs execute instructions in-order, so you're limited to instructions that are from the same basic block, as opposed to new x86s (K6+, PII+), which execute out-of-order, and can thus do at the same time operations from different basic blocks.

Now, the problem is that automatic vectorization works well-enough only if your language is Fortran (which has no dynamic memory). I guess that's ok, since a lot of libraries are in Fortran. But the language itself is grotesque (after all, it's the first real programming language), and a lot of people just write their stuff from scratch.

At the same time, manual vectorization can make your code completely unreadable.

Hasn't this been done before?

[wayne606]

Isn't anybody leasing CPU time on, say, amd64 machines? Somebody asked me where they could get that kind of time but I didn't know...

The price of $1/hr seems high, but let's say you want to do it yourself...

Buy 100 amd64 boxes - say $3K each after you factor in racks, switches, etc. $300K

Pay 2 staff people full-time to administer $200K/year

Rent space, pay for power, etc - $2K rent/mo + $1/kwh, 10KW/hr = $10/hr = ~$2000/mo = $48K/year

Over 3 years (useful life of the cluster) this adds up to more than $1M.

However there are 26280 hours in 3 years so you would pay $2.6M to rent these hours from Sun (of course, sparc != opteron either in price or power, but...)

So if you expect to keep your cluster more than 38% utilized you win by doing it youself. That is, assuming you have the money up-front...

The maths for this....

[Handpaper]

has already been done. Circumstances vary, but this will be useful for some, less useful for others.

Is it right for you? I have no idea, but I know someone who can advise you :)

lateral Prior Art (in case of patent offensives)

[Bluedove]

I heartily hope that what Sun (and IBM) are doing with renting grid computing time takes off. I believe it to be a good idea and a useful service. However, if somebody tries to patent it (or a business model), please refer to the public-released idea at http://www.halfbakery.com/idea/Distributed_20Compu ting_20Business_20Model#1057251600 to see if this idea from 2002 can be used to invalidate the patent.

Funny...now you can buy "grid computing time" on a network of computers. Back in the day, you could buy "computing time" on a powerful mainframe. The more things change, the more they stay the same.

They are not the first to market

[CEODRI]

DigitalRibbon.com is doing this exact same thing. we have priced out our contracts as low as .35 Cents for 700,000 CPU hours with a 16 Terabyte bata back up. We also have resolved issues of pricing to performance.

Re:It's too expensive. Paradigm shift

[CEODRI]

This is the new Paradigm Shift in how computing is utilized. Digitalribbon.com is already doing it. http://digitalribbon.com/constellation.html Necroman is also correct infiniband is very expensive!

Maybe Gateway had the right idea

[api]

"Build it and they will come" is attractive if you know they'll come but maybe Gateway had it right: rent out time on idling in-store demo computers. http://www.extremetech.com/article2/0,1558,1125371 ,00.asp

Let's see... Over 100 Apple stores, Xgrid http://www.apple.com/acg/xgrid/ iPod Linux...

Pixar may have a new place to render movies. :)

MD

CPU costs Vs. Researcher Costs vs. Blade Servers

[billstewart]

Dude, you're gonna get a Beowulf cluster of Dells :-)

There are 8760 hours in a year, so Sun would charge $8760 per CPU-year, which seems like a lot - you can buy Pentium white-box computers for about $200-300, and probably AMD64 for around $500 with minimalist video and CPU more comparable to a Sun. Call it $876, just to be a round 10% of Sun's yearly rental, and a lot of technician time per box. Alternatively, blade servers are going to cost you more per CPU, but be a lot easier to manage, and they're still easily under 20% of Sun's price. Obviously if you need to run one problem on 1000 CPUs for a couple of days, renting time makes sense, and if you need to run big-CPU problems all year, buying it's likely to make sense.

Not counting labor and real estate, the tradeoff's probably about a month or two of usage - a university can almost certainly cost-justify building their own, because they'll find ways to keep it busy. A business might or might not, and the real advantage for a business is that they get answers back in days instead of months.

Re:Portability - Mobility

[Doc Ruby]

Moderation -1
100% Flamebait

I explain the business value of the subject of the story, amidst widespread Sun-bashing posts. And express hope that the tech will improve computing. That's "Flamebait"? TrollMods fear the Sun.

Fluid Dynamics computing, or Weather forecasting.

[MDKKnD]

OK, it's not just about CPU power, it's also about vast amounts of ram, and time critical applications. A lot of people are saying it would be easier to run their program on a single PC over 41 days than spend $10,000 on the Sun cluster for an hour. But sometimes the programs Sun is targeting would just not work practically on a single PC, or even a small cluster. Take the 3D Animation Industry for example. One of the hardest programs to run are Fluid Dynamics software as it requires a very long time to run, and also requires excessive amounts of Ram. I have used Aura and Glu3D and they will happily use up 2 or 3 Gbs of Ram (the max Windows allows to a single process), and it will take over 24 hours to create a few hundred frames of animation. If you are a high end Visual Effects studio tasked to create the latest effects on a blockbuster movie, you can't wait months for the computers to sit there and think. They have to turn over shots daily. In that case, you will happily use a local super computer, or Grid computing array.