Towards Automatic Actor Pinning on Multi-Core Architectures ∗

To use multi-core processors to the best of their potential, many ap-plications should be rewritten so that they can benefit from the sev-eral available processing units. However, to optimize performance, the simple implementation of several parallel lines of execution in the application might not be enough. Modern processors possess a hierarchical architecture of memory, thus making the performance of an application dependent on the placement of the application threads. Furthermore, some applications that employ thread-to-core pinning often sacrifice their portability for the sake of performance. The actor model, available in languages such as Erlang and Scala, is an attempt to ease the development of parallel applications while at the same time shielding the developer from the underlying plat-form. In this model, the distinct parallel lines of execution of an application, or simply actors, interact with each other through the exchange of asynchronous messages. In such a setting, the place-ment of each actor during the execution of the application has con-siderable influence on the application performance. In this paper we describe a work in progress that aims to create a dynamically adaptive actor placement engine. Using application profiling in as-sociation with hardware counters, it will pin and migrate actors to processing units aiming to optimize performance. This will be done in an on-line fashion that is both transparent and automatic, there-fore not sacrificing portability