networkworld.com
But you're right, if there are any interdependencies then the interconnect becomes important.
I wonder how they'd count Folding@Home, then: 500K active clients, 6M total clients, but only a fraction of their clients are active at any given point in time.
When one side of the world is spiking, the other side is sleeping soundly.
EC2 sprang from the problem that Amazon had to buy a bunch of servers to handle the load around the holidays and these servers went underutilized during the rest of the year. So they decided to lease those resources.
When asked about what happens to EC2 during the holidays, the engineer basically replied that Amazon has priority.
(Will the first sign of a runaway AI be skyrocketing AWS spot prices?)
Statistically, adding multiple standard derivations yield a smaller value than their sum.
I guess you might be able to build a system in the cloud to provide TOP500 level of performance, but it would be pretty hard even with the fancy EC2 HPC instances (http://aws.amazon.com/ec2/hpc-applications/).
Although they do not provide an answer, here are some links to additional info - I spent some time searching for additional info on the Top500 setup, but found little:
* http://aws.typepad.com/aws/2010/07/the-new-amazon-ec2-instan... * http://news.ycombinator.com/item?id=1904590
In my experience Amazon did a pretty good job setting things up. It's fun to play around with HPC instances, you can get some sweet performance.
That's probably the only type of process that would work in the cloud. Most HPC applications require lots of communication between nodes, so I don't think I would call this a proper supercomputer.
This makes me believe someone is lying about something in this article.
Then, there are many problems out there which are not NP complete for which we are nowhere near to finding fast, accurate solutions. The problem is not that logic prevents us, but that we're simply not clever enough yet.
What I'm trying to say in a roundabout way is that spinning up many cores will not help you find a perfect, fast solution to an NP complete problem. And just because you have 10,000 cores that is not an indicator of it being difficult or hard to solve a given problem, regardless of its complexity class.
Example: a brute-force attack of an encryption algorithm that uses a 256-bit-key, would require trying out all possible keys, which is 2^256 ... which right now it would take far longer than the age of the universe to complete.
AND, most importantly, dividing that number by 10,000 (the number of computers in the article), or heck, let's be generous and say we have 1,000,000,000 computers ... would be absolutely meaningless.
It's simple really -- 2^256 / 1 billion computers =~ 2^226 -- and computing it still takes far longer than the age of our universe.
And lets say that with technology advances, you can have 70,000,000,000 computers (that's 70 billion computers, or a 700,000,000 % increase from the number in our article). Nevermind the energy required to power them or the storage capacity needed, or other such none-sense. So instead of 2^226, you now have 2^220 cycles to go through, an absolutely meaningless decrease, and still takes far longer than the age of our universe.
As a fun exercise, try figuring out how many computers would be required to bring that number down to ~ 2^200 -- that would still take far longer than the age of our universe to compute ;)
time for me to teach you something : http://en.wikipedia.org/wiki/Humour
""". Genentech benefited from the high number of cores because its calculations were "embarrassingly parallel," with no communication between nodes, so performance stats "scaled linearly with the number of cores," Corn said."""
That is a direct quote from the article.