Well I must admit that zip code was best case under ideal conditions ;-) There are always plenty of exceptions that put sand in the gears. Putting on my conservative hat the approach is more granular than guessing the right country as was being discussed before. My intention was only to infer there is more than one way to approach the problem and an approach that can avoid some of the DHCP issues seen in the datbase approaches. This was work from 6 years ago so a bit fuzzy on the particulars at this point. best, sean ----- Original Message ----- From: "Steven M. Bellovin" <smb@cs.columbia.edu> Date: Monday, May 15, 2006 10:25 pm Subject: Re: Geo location to IP mapping
On Mon, 15 May 2006 21:49:31 -0400, Marshall Eubanks <tme@multicasttech.com> wrote:
I seriously doubt this would work to better than the regional area.
My zip code (20124) region is about 5 km across, which would be
15
microseconds in vacuum, and maybe at most 50 micro seconds in glass. So, you would need accuracies at the 10's of microsecond level to specify zip codes.
I can believe that you can measure transmission times down a fiber and achieve repeatability at the microsecond level - in fact, I remember a Michelson interferometer that they set up at JPL / Goldstone that tested the Sagnac effect in glass, which required substantially better repeatibility than that.
But do you really think that you can estimate the router delay on the (for example) 9 hops between here and GMU to better than 1 millisecond each ? (That would imply a 3 millisecond rms error if these errors were random and Gaussian, or about 1000 km in vacuum, and maybe 500 km error in glass.)
So, I think that this would fail by at least 2 orders of magnitude for zip codes in a real operational network. Which coast of the US, sure, but not much better than that.
I suspect you can do that; a bigger factor is the link type of the lasthop. Cable modems, DSL, 802.11 -- they all have characteristic delays.
The important insight is that you care about *minimum* time. You can lots of queueing delays and jitter most of the time, as long as you get one packet through unobstructed. Send enough probes and you'll make it.
I did some similar work in 1992; see http://www.cs.columbia.edu/~smb/papers/netmeas.pdf for details. You couldn't repeat, today, exactly what I did then, because of the way pings are handled by modern routers, but I suspect one could find analogous schemes. To give one example of what I could tell -- and I was looking at the per-byte cost -- I was able to determine, from New Jersey, that a router outside Chicago was misconfigured; the site's backbone Ethernet should have been on the same card as the serial line (in the days of T-1 interfaces...), because copying the packet across the backplane introduceda noticeable per-byte delay.
--Steven M. Bellovin, http://www.cs.columbia.edu/~smb