Dynamic data reconfiguration for SPMD programs in faulty multicomputers

The single-program multiple-data (SPMD) paradigm is becoming the most diffuse way to program commercial multicomputers. In this paper we demonstrate that for a wide class of SPMD algorithms it is possible to achieve an efficient fault tolerance avoiding hardware redundancy. We propose a software approach that aims to reconfigure data, thus achieving a good slowdown in computation owing to the fine granularity of the workload to redistribute. In particular, we present and compare three data reconfiguration strategies applied to a problem model that includes a wide class of SPMD iterative algorithms characterized by nonlocal communications among the nodes. The result is that in most of the cases it is better to introduce some communication overhead than to leave idle a few healthy processors