High performance computing in the limelight

Businessweek has given high performance computing some coverage lately. One story talks about supercomputing for everyone. The piece begins with Dan Reed’s recent move to Microsoft as the company seeks to understand where supercomputing is going and how they can leverage it.

In a second story, a cover story on Google, they talk about Google’s grand ambitions for cloud computing, a far more sexy topic.

I have covered the cloud a lot lately, so lets talk about story #1 (found via InsideHPC). You have heard me moan about the lack of love for high performance computing in the media (with the notable exception of cloud computing). I’ve also grumbled about how the power of computers isn’t really being maximized and about how computers haven’t really taken that next leap. The first BW piece on supercomputing comes back to what Jack Dongarra said at PARA ’06 – “The problem is the software“. In other words, one of the challenges of tapping into all the processing power available to us is the lack of software that really maximizes all those cores and architectures.

However, the piece is somewhat schizophrenic and doesn’t really get to the core of the problem. First of all how do you define supercomputing? Assuming you just mean raw processing power, then to my mind supercomputing is a moving target. Yesterday’s supercomputer is my mobile phone (OK perhaps too extreme, but you get my drift). Supercomputing in the modern sense fits into two categories

1. Specialized applications – Protein Folding, Weather Prediction, etc.

2. Consumer/mass applications – Here I think of stuff like gaming, search, etc.

Specialized apps are are always going to be the realm of experts. All that needs to change is our access to resources. Shouldn’t anyone be able to tap into supercomputing resources with the appropriate access controls? We could have a specialized cloud where projects or users need to be approved so that resources are being utilized properly, perhaps a cloud of Blue Gene machines. Acclerated Computing would also fall into this category as acclerators are ideal for specialized high performance tasks

For consumer apps the question I have is this. For most of what we do on a daily basis do we need a “supercomputer”? As commodity processors get faster our ability to harness that power increases, and we’re always challenging the upper limit. Gaming, video production, etc are always pushing boundaries of consumer hardware. It would seem that others are asking the same questions, e.g. Justin Ratner, CTO of Intel.

Is this whole infatuation with performance something that has moved beyond what the vast majority of users really care about?

Are there really a set of applications that require 10, 100, 1,000 times the performance we have today?

if we have it at an attractive price point, will it drive high volumes?

The question, as he puts it, still remains. Is there a mass market for high performance computing? With our computers so fast now (too fast for most tasks) and an increasing amount of work being done on the cloud as opposed to the users computer, what is the market for HPC hardware?

That brings us to the cloud. IMO the cloud is where the market lies, and computer hardware vendors might need to start thinking of their supply chain differently. They are serving the consumer market with high performance hardware, but only indirectly.

Nick Carr thinks that Yahoo’s research chief is being a little premature and hubristic in stating that there are only five computers on earth; Google, Yahoo, Microsoft, IBM, and Amazon. While that is taking things a little too far (a lot too far), it is clear that the IT ecosystem is changing. In the coming years there will be the following

1. The Consumer Cloud – Amazon and Google will be the main players, with Yahoo and Microsoft likely to join in the fun
2. The Corporate Cloud – IBM will rule the roost here, but I suspect HP and Sun will get in on the action
3. Specialized HPC – Climate modeling, scientific computing, etc will fall into this category for certain tasks. The key here is understanding which type of resource a particular task belongs to and choosing resources appropriately

That leaves consumer hardware. If my thesis on the World Wide Grid (what Nick calls the World Wide Computer) is correct, then the way we use our computational resources is going to change radically, from software distribution, hardware pricing, etc.

This is a very US centric view, but you get my drift. From where I sit, it’s good to see computing getting love in the MSM.

In conclusion, one is left with a lot of questions. Just reading the two pieces, and even my own writings, and you can see that the industry is struggling as computing, even high performance computing, becomes increasingly commoditized, and user behavior evolves. How all the advances we’ve made in chip design and architecture will be put into practice is conjecture at this point. To some extent, if the cloud is where the action is, the user doesn’t really care. They just want their work done. My personal interest is to see the impact of the cloud and supercomputing on science. Historically, it’s tended to be difficult to get access to vast, scalable resources of compute power. Is that going to change? Will our software be up to the task? Science has done, by and large, a terrible job of tapping into recent advances, and allowing more people to access more resources inexpensively.

Technorati Tags: BusinessWeek, High Performance Computing