Adding Fault-tolerant Transaction Processing to LINDA

To simplify the difficult task of writing fault-tolerant parallel software, we implemented extensions to the basic functionality of the LINDA or tuple-space programming model. Our approach implements a mechanism of transaction processing to ensure that tuples are properly handled in the event of a node or communications failure. If a process retrieving a tuple fails to complete processing or a tuple posting or retrieval message is lost, the system is automatically rolled back to a previous stable state. Processing failures and lost messages are detected by time-out alarms. Roll-back is accomplished by reposting pertinent tuples. Intermediate tuples produced during partial processing are not committed or made available until a process completes. In the absence of faults, system overhead is low. The fault-tolerance mechanism is implemented at the system level and requires little programmer effort or expertise. Two implementations of the model are discussed, one using a UNIX network of workstations and one using a Transputer network. Data measuring model overhead and some aspects of system performance in the presence of faults is presented for an example system.