Google at Carson’s Speed
cringely.com
cringely.com
Apparently Google determined that they were spending more on electricity than on silicon. And in their massively parallel world, 3 processors running at 1/3 speed are just as good as 1 running at full speed. So they buy more machines, underpower them, and get the same results at lower cost.
As a result, Cringely claims, they have more computing power than they are using, and they think about what to do with that extra power. But I reply: they only have excess cycles if they switch back to running their servers at higher power. And they've already decided not to do that. Thus, no excess cycles, no unintended consequences, no "lead foot".
Right? Am I missing something? Is Cringely?
I was hoping that someone else would set the mess straight, including sorting out the vague definitions, but no such lazyweb luck this time ...
First thing that looks wrong to me is that server power estimate at one-third. I occasionally see CPU reviews with white-boxes benchmarking at something like 120W for idle power and 200W for full-throttle work, for CPUs rated at 90W or so. Perhaps racked servers have less overhead, but I doubt they do any significant work at 1/3 of full power. But I'd be happy to be contradicted by someone with real info.
P.S. Coincidentally, in the LWN piece I also linked yesterday (http://news.ycombinator.com/item?id=2564007 ), commenters samth and incase provided links to info about an IBM Research tool (Performance Analysis of Idle Programs, by Erik Altman et al.). Web page: https://researcher.ibm.com/researcher/view_project.php?id=13... . PDF paper: https://researcher.ibm.com/researcher/files/us-sjfink/res000...
EDIT: Carson's speed is clearly a real concept. However, a web search turns up nothing about the Breguet Number. Are we sure Cringely isn't just making this up?
Also, FWIW, my original point (see GP by me) is also made in the comment by NoTalentHack on Cringely's site.