Section: New Results

Empirical study of Amdahl's law on multicore processors

Participants : Carsten Bruns, Sid Touati.

Since many years, we observe a shift from classical multiprocessor systems to multicores, which tightly integrate multiple CPU cores on a single die or package. This shift does not modify the fundamentals of parallel programming, but makes harder the understanding and the tuning of the performances of parallel applications. Multicores technology leads to sharing of microarchitectural resources between the individual cores, which Abel et al. classified in storage and bandwidth resources. In this research report [39], we empirically analyze the effects of such sharing on program performance, through repeatable experiments. We show that they can dominate scaling behavior, besides the effects described by Amdahl's law and synchronization or communication considerations. In addition to the classification of Abel et al., we view the physical temperature and power budget also as a shared resource. It is a very important factor for performance nowadays, since DVFS over a wide range is needed to meet these constraints in multicores. Furthermore, we demonstrate that resource sharing not just leads a flat speedup curve with increasing thread count but can even cause slowdowns. Last, we propose a formal modeling of the performances to allow deeper analysis. Our work aims to gain a better understanding of performance limiting factors in high performance multicores, it shall serve as basis to avoid them and to find solutions to tune the parallel applications.