] > Isn't IBM doing some sort of fancy load redistribution for the WWW
] > servers its running for the 1996 Summer Olympics? I seem to recall they
] > were determining the "closest" server via a technique called "ping
] > triangulation", whatever that is.
]
] If so, that's pretty silly and is probably not leading to performance
] that's very different from either purely random or strict round robin.
Agreed.
] Ping times only correlate to available bandwidth when congestion is bad.
] Otherwise these two factors are militantly unrelated to each other.
Agreed very much. However, ping times can be used as latency
determination metrics. It may be perhaps that some really sexy
dns server pings the dns query originator, and picks up NAP
transit points en route, returning the appropriately close-ish
web server as an A record.
Triangulation implies that they are sampling ping times from two or more
different routes. They may be relying on this to warn them when paths are
becoming congested as opposed to using ping times as a means of measuring
which is the better of two uncongested paths. If the ping triangulation is
a process separate from web serving that is continuously sampling paths
and adjusting web server routing (redirects) dynamically then it may well
have some value.
They use the ping times to figure out which server would be closest.
All the servers are not located in the same place. The idea is that
european users may receive better service from a european server.
This brings to mind a question: are ping times a more appropriate vector
than hopcount or topological locality? Ping times reflect a lot of
important (but ephemeral) aspects of performance which more direct
measurements do not. E.g., the latency of trans-pond links nicely
reflects their cost in a matter not easily captured in simple topology
maps. Ditto for congested links which might be closer to the viewer.
Of course cacheing solves all of these problems (J <- hook next to bait),
but in this imperfect of worlds, what reasons, if any, make ping time
less attractive than other metrics? I used to think them simple-minded
and sloppy, but now I am not so sure.
And latency is disproportionately important when you're concerned only with
HTTP performance.
However, pings *can* be used to measure bandwidth, by sending multiple
sizes of ping and measuring the difference in response times between small
and large pings. On a graph of ping size vs. response time, you'd see the
y-intercept representing base latency and the slope representing bandwidth.
This is how bing does its work.