Hi Jim.
"I've often wondered about the efficiency of network power utilization, and how much power the 'net consumes."
I've asked similar questions here in the past. Qualifying which entities and elements to include as components of "the 'Net" is a formidable task, itself. Do we include the hundred-thousand plus telephone central offices and cable tv head ends? Their repeater sites and field electronics, too? There are hundreds of millions of residential PCs, modems, and CPE that attach to the 'Net. Certainly Internet Exchange Points (IXPs), colocation centers, network access points and host sites fit the bill. Individual content providers' data centers would, too.
Also, most educational-, enterprise-, research-, and government- data centers and storage sites, would be considered attachments to the 'Net, as well, as would the individual desktop devices and LAN electronics from each of those institutions be included. How about WiFi hotspots and Internet kiosks? EVDO and other 3G wireless formats?
One estimate that I read very recently - but cannot locate at the moment - attributed approximately 20% of the nation's power consumption to data centers alone. Again, what constitutes a data center, and which data centers qualify as being 'Net-dependent, or 'Net-attached?
Aside from the processing efficiency in servers, which certainly should be included as a means of gauging power efficiency, there is also the matter of how power is distributed within data centers to individual racks and servers that is gaining a lot of attention. Servers today are individually fed by dually-redundant (sometimes tridundant) independent a.c. power circuits that must undergo rectification in power supplies. A data center that houses a thousand servers, say, must accommodate, at a minimum, two-thousand a.c. to d.c. conversions, typically at between 65% and 75% efficiency, depending on their loads at any instnat in time. Another factor directly related to inefficient power consumption is the amount of heat that is thrown off, and the amount of additional air conditioning that is necessitated.
A more efficient means of powering servers, some argue, would be to create a central d.c. power plant and distribute low voltage d.c. (e.g., 5v and 12v power feeders) directly to each rack and/or shelf requiring it, thereby reducing the total number of inneficient conversions that must take place to a bare few. If efficiencies could be boosted through this means to ~ 80% on a sustained basis, say, then it could wind up saving Google alone some ten to fifteen million dollars per annum, if your earlier reference of $100Million was correct. But equally important would be the level of conservation that could be achieved.
As for optical, huge amounts of power are consumed in WDM and OADM gear, but their capacity yields are astronomically higher than their earlier copper counterparts. I don't have any benchmarks to offer in this regard, although I should, but I shall do some digging around and report back here if I find anything that sounds reasonable, and would hope that you and/or others would do likewise.
FAC |
