Current architectures of multicore machines are becoming increasingly complex due to hierarchical designs. Consequently, to achieve better performance stability, reproducibility and predictability requires a deep understanding of the interactions between multi-threaded applications and the underlying hardware. In this thesis, we study two important aspects for the performance of multi-threaded applications. We show that performance stability is an important criteria to consider in the process of performance evaluation, and thread placement is an effective technique in termes of program performance stability and improvement. We first study the variability of program execution times, defining a rigourous performance evaluation protocol, and analysing the reasons of such variability and its implications for program performance measurement. Then, we study the relation between the inter-thread data sharing and thread placement strategies on hierarchical machines. We consider various strategies where the same placement is applied for the whole execution of the program. While some of them rely on the characteristics of the application, others are not. We also present other thread placement strategies that allow thread migrations in order to exploit data sharing during different program phases.
