Matthew Toseland wrote: > Okay, so are we agreed on: > - A pDrop of 5%, determined at each request/retry. > - Some form of per-node failure tables (requires modifications to ULPRs). > - Fatal DNFs.
All sound good. > - On an RNF or an RO, we toss the coin again. I don't think that's safe. Moving on if we get an RO or RL is fine, even without tossing the coin again, but an RNF has visited some nodes (potentially a lot of nodes). We shouldn't ever branch with probability greater than pDrop, otherwise the average branching factor could be greater than 1 under some circumstances, at which point the end times will be upon us. In the case of RNFs, if we branch with probability 1-pDrop rather than pDrop as it sounds like you're suggesting, then we only need to get RNFs for pDrop/(1-pDrop) = 5.2% of requests on average to cause the sky to fall and fish to speak in tongues. (If that's not what you're suggesting then sorry for misunderstanding.) > - Predictable mean visited nodes of 1/pDrop. Only if we never branch, ie RNFs and DNFs are fatal (which might be the simplest solution). Cheers, Michael
