Fault-tolerant parallel programming with atomic actions

The Pact (parallel actions) parallel programming environment provides an easy-to-use parallel execution and synchronization model based on task parallelization. To give the programmer an abstraction for global data (even on distributed memory machines) the Pact runtime system uses virtual shared memory. Execution´s efficiency is improved with data-dependent dynamic load balancing and latency-masking by multithreaded servers. Fault tolerance in Pact is based on atomic actions and is guaranteed by the runtime system in a fully user-transparent way. This article describes the Pact runtime system´s design together with its logging and recovery algorithms for an implementation on a massively parallel distributed memory computer