Jesús Adán Flores Contreras

Abstract. Performance prediction methods applied to multiprocessor systems have been studied to estimate resource provisioning, workload management and for mapping tasks to the system resources in an efficient way. Several models have been proposed to achieve this. However, most of these models were designed to be used either on a specific platform configuration or specific application, which make them not suitable to use them in different contexts. Therefore, there is a need for platform- and application-independent models that can be used on different environments. To design such kind of models it is necessary to determine which resources have more influence on different kind of applications and how these resources behave during the applications' execution. Cluster and Grid systems are multiprocessors systems, which nowadays involve multi-core processors. Some of these system are composed of heterogeneous and geographically disperse resources, this fact makes more complicated to analyze their behavior and generate an abstract representation of the resources in order to represent them in a model. This research is focused on designing a platform- and application-independent performance prediction model. We focus on long-running scientific applications. The model should be able to deal with different long-running applications executed on different platform configurations (e.g. different CPU speed and amount of memory but involving the same hardware architecture).