Slashdot Mirror
Lights On But No One Home At Sun Grid
cygnusx writes "The Register reveals that Sun's pay-for-use grid computing services hasn't picked up a single customer yet." From the article: "The missing customers prove quite shocking when you consider that utility computing users must agree to be named in marketing programs as part of their contract with Sun - a fact learned by The Register and confirmed by a Sun spokeswoman. More than one year since it first started hyping the 'pay-for-use grid computing services' Sun is still weeks away from presenting a customer to the public. The program has proved much tougher to sell that Sun ever imagined."
I sent an email inquery to them right after it was announced, and no one ever contacted me. I even talked to someone at Sun (a different division), and still never heard from them....
Personally I think the idea might work, but it might not in this incarnation. There seems to be a fair chance that Sun can claim to be ahead of its time again, which has in some ways been a while. Which is a good thing in itself, Sun has historically been a nice company to work with but has suffered from some stagnation for a number of years.
When you develop a project like this, you usually need to have enough money to run it profit-free (and even revenue-free) for some period of time so you don't need to have customers commit to buying vapor from you in order for the thing to work.
That being said, however, you still need to have done some realistic market research. They should have at least contacted some friendly organizations (current customers mainly) to gauge actual interest in this thing beyond just the "that sounds cool" stage. The larger the financial risk involved in the project, the more market research needs to be conducted to mitigate that risk.
It sounds almost like someone at Sun got a "really cool" idea, and everyone else at Sun thought it was super cool too, and no one bothered to ask anyone on the outside. Or if they did, they only paid attention to the ones that said it was cool and ignored the others. Or they only asked people if it was cool, and never asked them if they would buy it if it were available.
It seems like Sun badly misread the market here, and I would assume someone in their marketing department is going to have a very bad day in the near future.
I'm on OpenBSD with KDE and when I click on the button to register for an account, nothing happens. There's the reason, bad programming. They don't have customers because people can't register! JERKS
Couldn't a company work something out with BOINC where they pay BOINC $.01 per CPU-hour, and $.01 per milestone to each participating member?
I bet people would sign up in droves if they could earn a little money for their free computer cycles. It could be paid quarterly or monthly using an online payment service like PayPal or through good old fashioned checks in the mail.
Just an idea, and for only $.02 per hour instead of $1.00 per hour.
Dave
I know why they don't get any science customers from my own experience. Basically, if you buy a cluster of your grant, you pay just for the hardware, everything else: electricity, cooling, network, comp support comes from the department's budget. These costs are not negligible.
If you tried to buy time from Sun, then everything goes from your budget... So, for an average scientist, who might be interested it is much cheaper to buy my own little cluster and piggyback on department's infrastructure...
SUN seems to have missed the mark. Even in research, budgets are an issue. If an organization has n dollars to spend in solving a problem, results are expected using n dollars. When you purchase equipment, you can re-compile /re-run the job without adding to your expenses. In order to properly budget a research task (in terms of CPU time) you must know the amount of time required to solve the problem. If you are responsible for the decision to use the SUN Grid and your problem takes longer to run than expected (even at $1/CPU/hour) you could get intro trouble.
For example:
A problem that runs 25% longer than expected
Budgeted: 10,000 CPUS x 16 hours = $160,000
Actual: 10,000 CPUS x 20 hours = $200,000
This 4-hour / $40,000 overrun could prove detrimental to one's career.
Find free books.
It's a small market, but not nonexistant. Anyone doing high-energy physics needs as much processing power as they can get. Companies doing genetics research (say, researching gene therapy) tend to need lots of compute time doing massive searches and comparisions of genetic databases. Insurance companies doing simulations and analysis need massive computing power. Special effects companies chew through computer time.
There is no question that massive amounts of compute power are needed. The question is: is it actually cheaper to rent the CPU time instead of just buying and managing the machines themselves? I'm less certain on that. While someone else has to worry about buying and maintaining the machines, you need to modify your workload to work on machines you don't control. The remote site may upgrade to an incompatible system to serve other customers. They could configure themselves to run whatever OS loadout you want, but that will cost more to setup and maintain. You typically need to send your workload across the public internet; putting gigabit ethernet between your cluster nodes so you can toss 2 gigabyte data sets around is relatievly ship. Getting a big enough network connection to set those datasets across the country is more expensive. Running over the internet is also more fragile. Oops, a backhoe just took out the connection. When something goes wrong, why does the provider care? Doing it in house means you have local staff you can lean on. A provider can be made to care, to provide guaranteed response times, but it'll cost you even more.
Search 2010 Gen Con events
They never asked who the heck would want this. Then they could refine things like costs and opportunities.
This was an idea which was not required. The cheapest thing in the world is a CPU cycle. Unless you're doing things that demand far more that a Beowolf clauter can deliver, like SETI, and that aren't proprietary, like no commercial products I know of, you don't actually WANT this service.
What where they thinking?
MSBPodcast.com The opinions expressed here are my own. If you don't like 'em... Think up your own stuff.
> It's a small market, but not nonexistant. ...
>Companies doing genetics research (say, researching gene therapy) tend to need
> lots of compute time doing massive searches and comparisions of genetic
> databases.
Except that's exactly the problem. You have this massive database. You can't just wave a magic wand and have the database appear on Sun's Grid. You have to upload it to them. Which takes, what, weeks? Years? Sure, maybe then their computation will be 10x faster, but you've lost the game from the start.
The performance comparision for the Sun Grid is:
Time to Upload Data + Time To Compute + Time to Download Data
vs
Local Time to Compute (since the data is already on your servers, presumably)
When you look at it this way, the Sun Grid becomes a lot less attractive.
Real grids solve this in various way, with data routing algorithms controlling where data is pushed out to -- but it works because your local computers are on your own grid with a high speed interconnect to the other players. Sun isn't part of your own Grid, so you can't employ these data routing strategies. And presumably Sun won't accept crates of DVDs in the mail and manually upload them to their server farm.
They're also in competition with the (free) academic grids.
It's a small market, but not nonexistant. Anyone doing high-energy physics needs as much processing power as they can get.
"As much as they can get" is very different from "huge numbers of intermittent cycles".
Academic phycists (like your link) aren't going to suddenly notice that they need 20,000 Ghz-hours within two weeks. If they had that kind of spurt demand, they'd be a good customer for Sun's grid.
Instead, users like that will have a fixed annual budget, and will try to maximize their Ghz-hours. Currently, you can get more computation per dollar by purchasing hardware, instead of leasing from Sun. That is unlikely to ever change, because Sun's pricing has to include overhead for the flexiblity of the grid, big bandwidth to move the data in and out, and liability/security costs to assure customers of their privacy.
Only if Sun achieves a superb economy-of-scale with administration will they ever be more affordable than in-house servers or other, simpler CPU rentals.
PS. Academics have a subsidized supply of white-collar labor in the form of grad-students, who can adequately admin the bought hardware for little cost.