Multiple networks for heterogeneous distributed applications

Abstract- We have experienced in our distributed applications that the network is the main limiting factor for performances on clusters. Indeed clusters are cheap and it is easier to add more nodes to extend the computing capacity than to switch to costly high performance networks. Consequently the developer should especially take care of communications and synchronizations in its application design. The FlowVR middleware offers a way to build distributed applications independently of a particular communication or synchronization scheme. This eases the design of distributed applications independently of their coupling and mapping on clusters. Moreover we propose a performance prediction model for FlowVR applications which is adapted to heterogeneous SMP clusters with multiple networks. In this paper we present an analysis of communication schemes based on our performance prediction model. We give some advices to the developer to optimize communications in its mappings. We also show how to use multiple networks on heterogeneous clusters to balance network load and decrease communication times. Since the FlowVR model is very close to underlying models of lots of distributed codes, our approach can be useful for all developers of such applications