From: Christian Wettergren <cwe@Csli.Stanford.EDU>

What I wonder is wheter the server congestion really showed that the protocol is flawed.

I never meant to say that the protocol was flawed in any way. I'm sorry if I gave this impression. (I used pretty much the same protocol on Hal1) Since I'm not the one who's writing the code, I will not try to tell you how it should be written, of course. My point was only that the central server approach does not scale. When we reach its limit (and it seems we have not reached it yet), we can use a hierarchical approach, and it is faster than the random one. But the random algorithm does have its strong points and we should not dismiss it out of hand. Maybe I got a little carried away in my advocating of the random algorithm. (another topic:) As for the updates to the client software, let me point out that I did 10 different versions of my own client when working on Hal1. Some machines worked for one week with version 1, while others needed many updates, due to different network and OS conditions. This is the main advantage of a well-defined (and stateless) protocol: it allows the server and clients to be all updated independently while the computation is running. -- Damien