Sun Grid Utility Goes Live for Employees

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

Comments

$1/CPU/hour?

[Anonymous Coward]

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

Re:$1/CPU/hour?

[dsginter]

Let's say that you have some number crunching that will take about 7000 CPU hours. Are you going to be happier waiting a year for your desktop to solve the problem or would you pay $7000 to get the answer in one hour?

Sun is betting that there are many people/businesses that fall into the latter category.

Re:$1/CPU/hour?

[TheSHAD0W]

You forgot the convenience factor. You'd have to wait a week, twiddling your thumbs, while that cluster ground away at the data. If there's a problem with the data, you may not find out until the end of the week, at which time your bosses will be pissed because you'll be telling them the projections will be delayed.

With Sun's service, you'll probably get the result within a few hours, not a week. If there's a problem the tests can be re-run with plenty of time before the presentation.

Of course, your bosses may be even more displeased about the extra $10,000 cost of the run than they would've been about another week's delay. Hope you talk fast!

Wrong scale

[jd]

CPU charging has been done within Universities for mainframes and super-computers for a long time. Usually, though, it'll be in terms of clock cycles used and priority. The former can be metered exactly, as the OS will have that information whenever it swaps the process in or out. The latter can be fixed at the start.

(For real-time processes, you can even fix the clock cycles in advance.)

The advantages of doing things on this scale are that most heavy tasks will take in the order of seconds - at worst, minutes - and it is unreasonable to charge people for absolute time when an unknown fraction is spent in process swapping, paging and I/O blocking.

A flat per-cpu rate won't work well, as people will simply write programs to hog the CPUs as much as possible, thus taking the least absolute time, thus resulting in the least cost. It will also reduce the timeslaces available to competitors, driving UP their costs.

In other words, it is a recipe for creating a gang-land in electronic form, where the roughest and most brutal coders will "win".

There is also no incentive for Sun to improve such an OS, either, as they profit from latency. An increase in latency increases the fees, so reducing efficiency is a source of income.

It is, however, entirely in line with their Networked Computer idea, which I can only guess they are trying to resurrect. (If you can rent CPU space, you don't need to buy a local CPU.)

bang for the buck

[cybergrunt69]

OK, so for one year of CPU time, maybe it is initially cheaper to buy a whitebox and install linux than it is to use this Sun solution.

However, if you got a linux whitebox to run this, not only would you have to worry about power costs, but also every other detail that comes with making sure your machine is running. What about patches, upgrades, network, bad hardware, runaway processes, general administration, backups, storage, etc? Most of the people here would be able to do the standard stuff that's needed, but I'm sure a business that needs "xyz" computed would gladly pay the 2x price. Not only would it do away with all the minor details, but they'd also have their results back in a significantly shorter amount of time! I'm too lazy to do math right now, but I'd say a year of cpu time could easily be done in less than month. That alone could be _the_ deciding factor and the justification for the expense.

Re:Who is going to use it?

[K8Fan]

What about 3D rendering? There are lots of people renting time on render farms right now to make deadlines. If I can run PRrenderman on these CPUs, and the price includes storage for my rendered frames, it might be price competitive to buying a lot of processing speed that I'm not going to be using 24/7.

Economics of Sun Grid Computing

[WillAffleckUW]

1. cost for a CPU
2. cost for the box for the CPU
3. cost for data storage
4. cost for monitors
5. cost for cooling for above 1-4
6. cost for power for above 1-5

Now, ask yourself, will the price of power go down if oil will cost $100 (current median bet in the Oil Futures stock simulation using real dollars as per WSJ)?

yes, there are ways to reduce those costs. but not everyone can.

Computing@home

[Saiyine]

What about releasing a grid client so every one could earn some bucks by letting their cpus work for others?

You know, like the spam bots in windows, but getting money!

--
Dreamhost superb hosting.
Kunowalls!!! Random sexy wallpapers.

Cool but what about licensing?

[adawg]

Sun wants us to run an app on their powerful grid, but what do software vendors think about us running their single lincense across, say, 100 CPUs?

So now, on top of the $1/CPU/hr, we have to buy a license for each of those CPUs.

Or else this will be very good for open source.

Re:No market there

[Cromac]

You joke, but MS Research has been working on building distributed computing into Windows for a while now. It probably won't be all that long until they either have a client, or something after Longhorn that will automatically distribute CPU load across a LAN.

http://www.betanews.com/article/Microsoft_Heats_Gr id_Iron_with_Bigtop/1104374194
http://www.microsoft-watch.com/article2/0,2180,174 6291,00.asp

It may be on the back burner at MS for now, but as we've seen many times if they perceive a market they're missing out on they can throw enormous resources at a project to get it to market.

Re:Not for big problems, then

[Space cowboy]

You know, I've got several linux boxes. I don't believe any of them cost $100 for every 20 days runtime! As for 250%, Oh boy! I have a bridge to sell you!

And, if it's my computer, what management are we talking about ? It's a program running on a computer. I start it. I wait. I analyse the results. What's to manage ?

If I buy 100 of these things, you use a simple batch script (I wrote one at college in about 2 days). Typing 'batch ' was all that was required to start something. Typing 'batch list' gave you a list of all running programs on the cluster, and typing 'batch progress []' gave you progress reports (written by the app into the file 'progress' in outputdir) for one, many, or all programs on the cluster. Similar commands for 'kill' 'stop' 'cont' etc. Easy. Our nodes were DECstations but I'm sure Linux is equally amenable...

My PhD was in Neural nets, we used to have nets that would take 20 hours or so to train, you would do that 10 times to make sure your results weren't anomalous, and then the input variables needed to be changed, and the whole thing re-run. This could quite happily take several weeks.

Now I can see that we could have rented time on the grid, but my point is that our use would categorise as continuously-required, and I don't think the sums add up.

So, assuming you huge problem is *really* parallelisable (lots aren't) you can split it over several hundred nodes and get an answer back relatively quickly. But most of us who need answers to such problems need to do this over and over again, so again I ask - where does the break-even point happen ? I suspect it's lower than Sun think it is...

Simon.

Re:Not for big problems, then

[museumpeace]

economics and scalability are not that simple to mix. Your $400 machine simply can't handle the dataset or the array sizes or the threads or much any metric you'd get when you size up any number of huge but very practical programs that businesses sometimes run. To name but one example: the monster linear systems that model supply chain and what-if a huge number of variables so the company can find the right product mix...the list of such programs is bigger than you may be aware.

Save money: procrastinate

[dazedNconfuzed]

There's a wonderful technical paper analyzing the value of procrastination. Don't recall the details, but the upshot was that for certain computing problems long enough (on the order of years compuatation time), at some point it's just better to do nothing at all until the price of the hardware drops and the CPU speeds rise enough so that you can actually get the job done sooner than if you started running the job on what you had ASAP.

If you have computing job that will take 4 years on today's processors, and processors will be 4 times faster 2 years from now, better to do nothing for 2 years and then run the job on computers that will complete the job in 1 year, getting you done a year early including a 2 year vacation.

Wish I still had that paper...

Re:Everything Old is New Again

[pdo400]

I don't know if you're a simple troll or a simple-ton, but either way your logic is utterly flawed.

I work producing a software product that I personally wouldn't use. Does that mean nobody would use it? Apparently not since my company is profitable and I get paid.

Also, 'backwards' is a matter of perspective. For ordinary home users who have trouble keeping their desktops clean and their start menus in order, don't keep backups of their data, and throw away their PCs every couple years because of viruses and spyware, a thin client solution would certainly be considered a step 'forward'.

Missing the point

[tsotha]

I don't see why so many people think this is aimed at people doing massively parallel processing. It's not.

I'm thinking of my own employer, who's got hardware up the wazoo that mostly just sits around heating the building. Take payroll, for instance. They probably run some payroll batch job a couple times a month, then the rest of the time the computer that does payroll is just sitting around. Sure, you could run other stuff on it, but then when it came time to do the payroll everything would run slow and people would be upset.

The way I see it, this is perfect for all kinds of periodic batch-type business applications where you really want to have a dedicated machine but it won't be utilized all the time.

Also, machines that mostly sit around have about the same maintennence requirements as heavily-used hardware. They still need OS patches, security patches, backups, etc. And they need to get modified when the new head of IT decides logs should go in /var/tmp/messages instead of /var/adm/messages or whatever. I know my employer could probably get rid of 2/3 of its data center and a corresponding fraction of the high-priced administrators currently making everything run smoothly.

You could argue companies can do the same thing by balancing existing hardware, i.e. have one box host multiple "bursty" applications so the CPU doesn't have much idle time. But that takes lots of effort to manage, and it doesn't leave you any spare capacity when you need it. This way you don't need any spare capacity, and when your business grows you never have to worry about running out of rack space or lead times for new hardware.

The downside, of course, is you're trusting Sun to have the capacity available when you need it. In a way, it really is like a utility - Sun will be hoping its customers "bursts" will all average out to some managable load. I wonder if it's true.