Not to harsh on the design at all, but is round trip time (the F5 solution) the best way to calculate which distributed hosting / caching site a client connection should go to? Variance in rtt from one trace to another, or variance in hop count from one trace to another, could lead to inconsistent balancing results as the client/server connection is going to be stuck in some kind of quasi-flow cache and additional requests forwarded from the client to the same server. Is it safer to use a bgp AS_PATH or MED over pure rtt? (granted, routes can flap and cause the same effect, but is it as much of a problem) Is anyone willing to share experiences with one or the other? Cheers. Travis ----- Original Message ----- From: Sean Donelan <sean@donelan.com> To: <nanog@merit.edu> Sent: Sunday, December 05, 1999 6:04 PM Subject: Re: Mars lander (Re: How to achieve application reliability)
On Sun, 05 December 1999, Nathan wrote:
Here is a great link explaining in detail the design and architecture of the web accelerators on the net that were designed to support the load
that
marspolarlander.com was supposed to generate.......although they have no feed to push....it was impressive.
Yes, it was very impressive. I was hoping they would be successfull, because in a few weeks the next event will also require some interesting web site hosting. If it had worked, then the people who follow could just copy it. Even without a feed to push, the sites were unresponsive for parts of the day.
http://www.newsbytes.com/pubNews/99/140268.html
I'm very concerned we still haven't proven we know how to handle big event sites on the Internet.