Adam Lackorzynski, TU Dresden, Germany
Benjamin Engel, TU Dresden, Germany
Marcus Völp, TU Dresden, Germany

Caches are a well known mechanism for improving energy consumption and average performance by keeping frequently-used data near processing resources. Yet, at the same time, they form a major hurdle for worst-case execution time analyses, in particular if they are shared between multiple cores. Exploiting that most shared data objects are accessed or at least committed in a mutually exclusive manner, we demonstrate, using an ARM MP-Core multicore system, an explicit write-back synchronization scheme for shared data. We will analyze the performance and overheads it incurs, stressing predictability aspects. To our surprise, neither the local nor the shared caches need to be kept coherent for this scheme to work.