Fault tolerance in parallel implementations of functional languages

It is suggested that fault tolerance at the computing-model level is desirable in multiprocessors and that computing models for inherently parallel functional language programs provide for implicit fault-tolerance through temporal and spatial redundancy. While both extensional and intensional computing models can achieve this, it is argued that intensional computing models are much more efficient in tolerating omission and corruption faults. It is shown that demand-driven implementations (instead of data-driven implementations) of the intensional computing model can naturally realize fault-tolerance. The implementation of this approach in a parallel software system based on an intensionally modeled language called GLU is described. It is noted that fault tolerance at the computing model level is transparent to both the parallel applications programmer and the parallel computer system architect.<>