On Apr 8, 2009, at 17:16 PM, Shawn Willden wrote: > This is a big part of the reason why I think the loss modeling work > is important. In order to get good performance, you want n to be > large. On a stable grid of small to moderate size, you probably > want to set m equal to the number of available servers, then adjust > n to trade reliability off against performance and expansion. > Higher values of n provide better performance (up *and* down) and > less expansion but worse reliability. > > This means that it's tempting to set n close to m. But if you do > that, just how bad *is* your reliability?
I agree with all of this. I have questioned some aspects of your modelling work, but I didn't intend to say that it was worthless! Only that it is tricky for you to produce a good model, and then it is *still* tricky after you've done that for someone else to understand how safe it is to use that model to manage real world decisions. Which I'm sure you already appreciated... Regards, Zooko _______________________________________________ tahoe-dev mailing list [email protected] http://allmydata.org/cgi-bin/mailman/listinfo/tahoe-dev
