Nonintrusive Snapshots Using Thin Slices

This paper gives an efficient algorithm for recording consistent snapshots of an asynchronous distributed system execution. The nonintrusive algorithm requires 6(

n

–1) control messages, where

n

is the number of processes. The algorithm has the following properties. (P1) The application messages do not require any changes, not even the use of timestamps. (P2) The application program requires no changes, and in particular, no inhibition is required. (P3) Any process can initiate the snapshot. (P4) The algorithm does not use the message history. A simple and elegant three-phase strategy of uncoordinated observation of local states is used to give a consistent distributed snapshot. Two versions of the algorithm are presented. The first version records consistent process states without requiring FIFO channels. The second version records process states and channel states consistently but requires FIFO channels. The algorithm also gives an efficient way to detect any stable property, which was an unsolved problem under assumptions (P1)-(P4).