Post a Comment
I'm intrigued by this idea - the IBM researchers seem to have identified a somewhat overlooked series of issues with regards to clusters of commodity systems. Now they have a proposed solution. With crazy ideas like portable data centres, floating data centres cropping up, and proposals of standardized DC power for PCs, it seems like IBM is on the right track by trying to address concerns of power/space use.
From what I understand of their idea, it could best be described as having the Internet run a supercomputer - not the other way around. The 'net, as a whole, is a huge number of interconnected systems passing messages by a wildly varying set of defined protocols. Up until now, supercomputers (in my understanding) use a constrained, high-performance protocol to attack single problems (like nuclear warhead decay, geological surveys, etc). The proposal here seems to explore the ramifications of having many protocols passing data between many processes with different tasks and responsibilities running on a supercomputer.
The article at the reg notes that BlueGene is already a form of hybrid massively-SMP/cluster system (having attributes of both). It looks like the boys at IBM could be en-route to some interesting results!
http://weather.ou.edu/~apw/projects/kittyhawk/kittyhawk.pdf
The Reg article exaggerates what Kittyhawk is really about; the actual paper is pretty interesting.
See "5 reaons it might not be â€Blue Skies†for IBM"
http://www.johnmwillis.com/other/5-reaons-it-might-not-be-blue-skie...





