PRISM: an intelligent adaptation of prefetch and SMT levels

Current microprocessors include hardware to optimize some specifics workloads. In general, these hardware knobs are set on a default configuration on the booting process of the machine. This default behavior cannot be beneficial for all types of workloads and they are not controlled by anyone but the end user, who needs to know what configuration is the best one for the workload running. Some of these knobs are: (1) the Simultaneous MultiThreading level, which specifies the number of threads that can run simultaneously on a physical CPU, and (2) the data prefetch engine, that manages the prefetches on memory. Parallel programming models are here to stay, and one programming model that succeed in allowing programmers to easily parallelize applications is Open Multi Processing (OMP). Also, the architecture of microprocessors is getting more complex that end users cannot afford to optimize their workloads for all the architectural details. These architectural knobs can help to increase performance but it is needed an automatic and adaptive system managing them. In this work we propose an independent library for OpenMP runtimes to increase performance up to 220% (14.7% on average) while reducing dynamic power consumption up to 13% (2% on average) on a real POWER8 processor