Heterogeneous architectures for computational intensive applications: A cost-effectiveness analysis

Current workstations can offer really amazing raw computational power, in the order of TFlops on a single machine equipped with multiple CPUs and accelerators, which means less than half a dollar for a GFlop. Such result can only be achieved with a massive parallelism of the computational devices, but unfortunately not every application is able to fully exploit them. In this paper we analyze the performances of some widely used, computational intensive, applications, like FFT, convolution and nn-body simulation, comparing the performance of a multi-core cluster node, with or without the contribution of GPUs. We aim to provide clear measure of the benefit of a heterogeneous architecture, in terms of time and cost, with a stress on the technology adopted at different levels of the software stack for the application parallelization.