Parallelization of an algorithm used to simulate Atomic Force Microscope images

Abstract

We present the parallelization of an algorithm used to simulate images obtained with an Atomic Force Microscope. We use the standard parallel programming environment PVM so that the parallelization be architecture independent. We propose different algorithms: with or without large transfers of data between master and slaves and with or without load balancing. We compare execution run times on three kinds of platform, namely networks of homogeneous or heterogeneous workstations, symmetric multiprocessors with shared memory and a massively parallel machine with distributed memory. For this specific problem, the use of PVM leads to a good speedup for all architectures. However, we also show that the best algorithm depends on the kind of architecture. In particular a load balancing strategy is important for efficiency on an heterogeneous platform, whereas it has only a small effect on the other platforms.