The lightlink supercomputer
Due to my close industry connections, several months ago I heard that local Ithaca NY internet service provider Lightlink planned to purchase a supercomputer and rent out processing time. It appears that I can rip up my NDA, so to speak, since there’s an article in the Ithaca Times this week about this project.
It seems that it’s more of a cluster than a supercomputer, though. There’s a lot of talk about the number of CPU’s (1000 to start). It’s probably a Beowulf (wikipedia) cluster. (Geeks — feel free to weigh in with speculation on hardware, software and/or performance in tthe comment section). This approach is suited to applications that are parallelizable in in which each execution thread is mostly accessing a small set of local memory — I know of applications where the memory required would be in the order of tens of GB, and the execution thread would be accessing this in an unpredictable manner, so all of the memory needs to be fast.
Google’s servers are mostly clusters of commodity PC’s — but Google’s applications parallelize very well [The famous Google IEEE paper, now getting a bit out of date].
I was going to say something about the energy cost of 1000 CPU’s — using the numbers in the google paper, it might be something like $120,000 per year. But that paper is a few years to of date, and nowadays there are 2 or 4 CPU’s on a single chip that uses the same amount of energy, so perhaps it could be a factor of 5 lower — perhaps $24,000 per year.
