Computational Machinery

20 years after Jim Gray and Gianfranco Putzolu describe the five minute rule as a breakeven point for the tradeoffs of holding data in RAM versus disk, new authors re-evaluate the claims with today's pricing and capacities. They find that the five minute rule still holds for ever increasing page sizes. The Five Minute Rule

A Facebook VP laments that new AMD and Intel processors aren't performing in reality the way marketing would have you believe. I personally believe that this makes total sense. Most of the work that Facebook does is limited by memory bandwidth when serving up data to users, or disk bandwidth when doing analytics for internal use. While the Core i7 does address the memory bandwidth issue to some degree, the fact is that RAM -- relative to processors -- is only getting slower. So while there might be some gains for a few of the things that Intel and AMD tune their microarchitecture for (namely SPEC) nobody has put together a benchmark for commodity compute clusters and as a result none of the vendors can tune for the use case. Our aim here is to slot right into the niche that companies like Facebook, Amazon and Yahoo have need of and to satisfy the demand. New Chips Don't Deliver

"The human brain requires about 25 watts of electricity to operate. Markram estimates that simulating the brain on a supercomputer with existing microchips would generate an annual electrical bill of about $3 billion." If you do the math it comes out to just over $5700 per minute. I probably don't eat $5700 worth of food in a year.

I spent some time working in a high performance computing lab during my graduate studies. The P.I. use to always say that there are two classes of algorithms: those that can be run, and those that cannot. For now the human brain is relegated to the latter. Some day it will no longer be and when that day comes, the cost of using it will be crucial. Simulating the human brain

SGI has gone and done something smart. They've brought the extremely low power, highly efficient Atom processor to bear in the supercomputing world. With lots of memory bandwidth per clock cycle they just might be on to somthing. SGI 'Molecule' in development

As always, a thought provoking look at the big picture in terms of where software is going these days. Bob thinks that SQL is on the way out and that unstructured databases are up-and-coming. SQL is the modern-day buggy whip

This runs along similar lines to what we're doing. But the processors are awfully anemic and I don't believe that you can get enough RAM per node to be all that worthwhile. They've certainly gone for smaller, less powerful, and greater numbers though. Carnegie Mellon creates low-power cloud device

Just got a link in the mail today about this. Amazon offering a Hadoop service without a whole mess of configuration nightmares sounds pretty good. Especially if it's not a core competency or you don't need to be 100% sure all the results are precisely correct. Amazon's announcement

A stereotypical first blog post. Needed to get things rolling at some point. I'll also include a link to one of the first companies I've discovered to be in roughly the same ballpark. Physicalized servers as an alternative to virtualization