分布式算法课件7.ppt

Round lower bound, f = 1 Next execution, α2, removes message 1 →3. α1 and α2 indistinguishable to everyone except 1 and 3, hence to some nonfaulty process. Next, remove message 1 →4. Indistinguishable to some nonfaulty process. * Continuing… Having removed all of process 1’s msgs, change 1’s input from 0 to 1. Looks the same to everyone else. We can’t just keep removing msgs, since we are allowed at most one failure in each execution. So, we continue by replacing missing messages, one at a time. Repeat with process 2, 3, and 4, eventually reach the last execution: all inputs 1, no failures. * Special Case: f = 2 Theorem 6: Suppose n ≥4. There is no n-process 2-fault stopping agreement algorithm in which nonfaulty processes always decide at the end of round 2. Proof: Suppose A exists. Construct another chain of execs, each with at most 2 failures. This time a bit longer and more complicated. * Special Case: f = 2 Proof: Suppose A exists. Construct another chain of execs, each with at most 2 failures. Start with α0:All processes have input 0, no failures, 2 rounds: Work toward αn, all 1’s, no failures. Each consecutive pair is indistinguishable to some nonfaulty process. Use intermediate execs αi, in which: Processes 1,…,i have initial value 1. Processes i+1,…,n have initial value 0. No failures. * Special Case: f = 2 Show how to connect α0 and α1. That is, change process 1’s initial value from 0 to 1. Other intermediate steps essentially the same. Start with α0, work toward killing p1 at the beginning, to change its initial value, by removing msgs. Then replace the messages, working back up to α1. Start by removing p1’s round 2 messages, one by one. Q: Continue by removing p1’s round 1 messages? No, because consecutive executions would not look the same to anyone: E.g., removing 1 →2 at round 1 allows p2 to tell everyone about the failure. * Special Case: f = 2 Distinguishable to everyone. So we must do something more elaborate. Recall that we can allow 2 proc