Low-overhead fault-tolerance support using disc programming model

Abstract

DISC is a newly proposed parallel programming paradigm that models many classes of iterative scientific applications through specification of a domain and interactions among domain elements. Accompanied with an associated runtime, it hides the details of inter-process communication and work partitioning (including partitioning in the presence of heterogeneous processing elements) from the programmers. In this paper, we show how these abstractions, particularly the concepts of compute-function and computation-space objects, can be also used to leverage low-overhead fault-tolerance support. While computation-space objects enable automated application level checkpointing, replicated execution of compute-functions helps detect soft errors with low overheads. Experimental results show the effectiveness of the proposed solutions.