| Popis: |
Many parallel applications do not scale as the number of threads increases, which means that using the maximum number of threads will not always deliver the best outcome in performance or energy consumption. Therefore, many works have already proposed tuning strategies, which can be online or offline, to optimize performance or energy. Online tuning approaches (i.e. at execution time) are the most efficient, since they can catch intrinsic characteristics that can be only known at run-time (e.g., input set, current load balance, microarchitecture details). However, such dynamic nature requires fast design space exploration, since online tuning will always add extra overhead to the execution. In this online process, this work investigates how parallel regions may influence each other during execution and shows that data migration events may represent a considerable overhead, affecting the tuning progress and impacting the execution time and energy consumption. Hence, we demonstrate why many approaches will very likely fail when applied to simulated environments or will hardly reach a near-optimum solution when executed in real hardware. |
