Sun Grid Compute Utility

jbltgz writes "The Register is reporting that the long awaited Sun Grid Compute Utility has been opened to the public. Now you can run your CPU intensive jobs on a grid of AMD Opteron-based Sun Hardware for $1 per CPU per hour for a fraction of cost, in a fraction of the time."

A bug could be costly

[DrDitto]

I use grid computing for simulations. If I were charged for CPU-hours, you can bet I would be more careful about debugging. I've wasted thousands of CPU hours because of bugs, or sloppy configuration, in my simulator generating incorrect results. One bug was an infinite loop that resulted in 100 CPUs spinning for a week before I noticed!

Re:Since I don't remember the earlier stories...

[AKAImBatman]

Do you (the customer) supply the software to run on these distributed boxen?

Yes. See the Sun FAQ [sun.com].

Cause if that's the case, I can see a business model that involves lophtcrack or John the Ripper.

Considering that Sun has your Credit Card number and your contact information, I highly doubt you'll be getting away with much DDOSing or Spamming. Even if you use a stolen card #, Sun would be likely to kick you off the servers after they notice excessive network and/or mail server usage.

Re:I would have made use of Sun's grid already

[shaitand]

You could save some really serious money by porting to a proper compiled language.

Re:The Sun is setting

[Anonymous Coward]

Cost of writing applications and porting them should be included in any estimate, yet you only include it in the estimate of Sun's cost and not your own. Also worth pointing out is that your "years old" hardware wont perform anywhere near the level at which suns offering will.

You also failed to mention anything about this "less than $500k" cluster. For the price of a $499k cluster (which again doesnt include man hours spent on code, so it wont be counted against sun either) I could rent 50 cpu's 24x7x365 x1. You get more cpu power buying a system at that price point. Now factor in maitenance costs. Failure costs. Power and cooling costs. Personell costs for admining the setup (not coding). You still get the edge buying the system, but by how much ? The reality is this is not geared towards computer people. Its geared towards math, graphics, and physics people. Groups who dont always have the time and resources to buy, setup and admin a cluster but who have some data once in a while that they need to proccess.

Now apply your math to someone who has something that needs to be done once in a while in a condensed time frame. This is one hell of a lot better system for that then buying a cluster and hoping it doesnt break right when you need it ... or while it sits idling for the majority of the year.

Re:The Sun is setting

[__aaahtg7394]

It gets even better when you consider that Sun's smaller Opteron kit starts at about 2k$/node. So, if you need more than a few months' processing, you can just buy the boxes and build the infrastructure for about the same cost.

I suspect the real selling points are:
- Sun's service is probably straightforward for app developers.
- The hardware is essentially "infinite."
- "Oh, you need a month's worth of processing done by next Monday? We'll have it done Saturday night, if you'd like to pick it up then, ma'am."
- Two words: volume pricing.

But, I have a hunch that the real point is marketing. "Oh, you need a 1,000 node grid? We happen to have expertise in building just such a thing. Here, give this information to your apps people, let them try it out, see if they like it. We can install just this setup for you here, if you'd like."

Re:Imagine All the SPAM!

[Wesley Felter]

That's why the grid nodes are not connected to the Internet.

Re:The Sun is setting

[Tsunayoshi]

The whole point of this was stated by John Schwartz when the idea first came up. This is moslty for the market of people who only do massive number crunching on a few occasions but often enough that they would maintain their own computing cluster. Now, these same people can get rid of the hardware, people, power, cooling, and lab space and the other costs associated with maintaing a cluster and just worry about paying for CPU time the few occasions they actually need it. And if suddenly you have an immediate need for 10x the computing power, let Sun know and you can get more systems involved for your project. If you maintained your own compute cluster, how easily could you add 50 cpus to a task if you are already running at 100% CPUS??

Condor

[Isaac-Lew]

Why wouldn't a mid- to large-sized organization use something like Condor [wisc.edu]? Just install it on everyone's (linux or win2k/xp) server/workstation, maybe set some prioritization scripts so that it would use more resources after-hours (when most people are out of the office, but have their systems on anyway), & save themselves $$ instead of paying to have their data on someone else's remote system?

Re:Details please

[dubiousdave]

It's not just a question of frequency of use. In order to have your own cluster to do this sort of work, you have to have a lot of infrastructure that many organizations can't really justify. A large datacenter can cost a fortune every year in direct power use by the computers, plus very heavy cooling requirements for the room. You also have to have one or more people running the cluster. Then, on top of all of that, you have to deal with upgrading and maintaining. It's a huge pain that a small, say, fabless semiconductor design firm, may not want to deal with. They focus on their core business of designing circuits and send those designs off to someone else's cluster to test.