Transparent optimization of OpenMP applications via thread throttling and boosting techniques

Abstract

The growing number of cores in modern multicore architectures has brought together the need for better use of hardware resources. Consequently, two techniques have become widely used to optimize the performance and energy consumption of these environments: Dynamic Concurrency Throttling (DCT) and Boosting. On the one hand, DCT adjusts the number of threads in parallel regions to minimize the effects of intrinsic characteristics of the applications that impact performance and energy consumption (e.g., data synchronization and communication). On the other hand, Boosting techniques focus on making the performance reach its maximum level during all phases of the application by increasing the processor frequencies while respecting the Thermal Design Power (TDP). One of the main challenges is that each region of a parallel application can behave differently (i.e., memory access behavior and degree of parallelism) which makes using both techniques combined not a straightforward task. Choosing the wrong number of threads and enabling/disabling boosting frequencies in the incorrect phases can lead to increasing the energy consumption and performance degradation. To solve this problem, this work presents two strategies that apply DCT and Boosting to improve the trade-off between performance and energy consumption (represented by the energy-delay product - EDP): PFG, a strategy that optimizes each region of a given application, individually; and PCG that considers the combination of parallel and sequential regions during optimization. Both strategies are transparent, automatic, and deeply integrated into the OpenMP parallel programming interface, so no code modification or recompilation is necessary. By executing twelve well-known benchmarks in three multicore systems, PFG and PCG improve EDP by up to, respectively, 95.3% and 95.5% compared to standard OpenMP execution, 90.9%, and 94.8% on Varuna-PM and 80.5% and 83.7% against the Core Packing technique. We also show that PFG is more suitable for applications with high variability in the CPU workload, while PCG is better when there is low workload variability.