Performance Implications by the Hierarchical Design of Clusters

Colloque avec actes et comité de lecture. internationale.International audienceWe present experimental results for the evaluation of PC clusters that differ on several aspect of their architecture, such as being mono or biprocessors and having a high bandwidth interconnection network or not. These experiments confirm that a good equilibrium between the speed (throughput) of the different components is crucial for a satisfactory performance of such a machine. On the other hand they also show that the latency of the underlying infrastructure is of less importance and can be hidden by applying techniques from coarse grained parallelism