Performance Prediction for Mappings of Distributed Applications on PC Clusters

Abstract

Abstract. Distributed applications running on clusters may be composed of several components with very different performance requirements. The FlowVR middleware allows the developer to deploy such applications and to define communication and synchronization schemes between components without modifying the code. While it eases the creation of mappings, FlowVR does not come with a performance model. Consequently the optimization of mappings is left to the developer's skills. But this task becomes difficult as the number of components and cluster nodes grow and even more complex if the cluster is composed of heterogeneous nodes and networks. In this paper we propose an approach to predict performance of FlowVR distributed applications given a mapping and a cluster. We also give some advice to the developer to create efficient mappings and to avoid configurations which may lead to unexpected performance. Since the FlowVR model is very close to underlying models of lots of distributed codes, our approach can be useful for all designers of such applications.

Mapping Heterogeneous Distributed Applications on Clusters

“…The performance of a FlowVR application can be described by a formal model, as introduced in [8]. This model takes several inputs :…”

“…We propose in [8] to study the application graph to determine implicit synchronizations between modules. Then, we compute the frequency F m of modules according to their execution times, implicit synchronizations and explicit ones defined with synchronizers.…”

“…Then, we compute the frequency F m of modules according to their execution times, implicit synchronizations and explicit ones defined with synchronizers. Note that the processor load required for each module may vary depending on synchronization between modules, thus, we provide in [8] an algorithm to determine for each module the new value of its load, noted Load c m . At the end of each iteration, each module m sends messages on its output ports, thus we can compute the required bandwidths bw l for each network link l from F m and V ol oi m .…”

“…The load Load c i,j put by a module M odule i on a processor p j is expressed in percents and is determined by our performance model [8]. A first restriction is that a processor cannot be loaded beyond 100%.…”

“…The search for a suitable allocation uses the performance model of FlowVR introduced in [8]. Some general constraints are defined from the FlowVR model and its performance model and are common to all FlowVR distributed applications.…”