Author(s): Smith J, Watson P
Abstract: It is argued that there is a significant class of pipelined large grain data flow computations whose wide area distribution and long running nature suggest a need for fault-tolerance, but for which existing approaches appear either costly or incomplete. This paper presents an approach which exploits limited input from the application layer to implement a low overhead recovery protocol for such data flow computations. Over a large range of possible data flow graphs, the protocol supports tolerance of a single machine failure, per execution of the computation, and in many cases a greater degree of fault-tolerance. The protocol is implemented within an emulation of a distributed query processing system. Preliminary performance measurements suggest that the overhead is indeed low.
Keywords: data flow, fault-tolerance, measurement, query processing, rollback-recovery, wide area
|
Professor Paul Watson
|
|