Preditor de Desempenho de GPUs aplicado à Exploração do Espaço de Projetos ciente de Dark Silicon

Abstract

Simuladores de sistemas heterogêneos GP-GPU procuram oferecer acurácia de desempenho ao custo de elevado tempo de execução. Com o objetivo de evitar o custoso processo de simulação durante as etapas de exploração arquitetural de sistemas baseados em GPUs, desenvolvemos e avaliamos diversos preditores de desempenho de GPUs baseados em algoritmos de aprendizado de máquina com acurácia e baixo custo computacional. A qualidade dos preditores desenvolvidos neste trabalho foi avaliada por meio de métricas como coeficiente de determinação, score de treinamento e validação cruzada. Preditores baseados nas técnicas de Random Forest e SVR apresentaram os melhores resultados tanto em acurácia quanto performance.

Performance predictors for graphics processing units applied to dark‐silicon‐aware design space exploration

“…The same authors of this work have already published some papers in the same context, ranging from the design of MultiExplorer, 3 the dark‐silicon estimates for CPU and GPU, 30 the integration of CPU performance predictors 34 and a preliminary version of the integration of predictors of GPU performance 35 …”

“…We replaced the optimistic MultiExplorer approach for performance estimates by a SVM-based regressor to CPU performance estimates systems 34 and a Decision Tree-based regressor to GPUs. 9,35 In order to explore heterogeneous architectural alternatives, we have set MultiExplorer DS-DSE module to use up to two types of computing cores: one is original computing unit (16-core Smithfield) and other alternative cores (any other from the Table 8), which must meet the area and power density constraints, ranked for the performance they can achieve.…”

“…The same authors of this work have already published some papers in the same context, ranging from the design of MultiExplorer, 3 the dark-silicon estimates for CPU and GPU, 30 the integration of CPU performance predictors 34 and a preliminary version of the integration of predictors of GPU performance. 35 Currently, the team's efforts are to extend the methodology to other accelerators such as FPGAs or ASICs and approximate computing. Regarding the extension to GPUs, it is still possible to carry out a more in-depth study on how to characterize the performance of GP-GPU systems, as the simulators used in the current paper are very limited to deal with this.…”

“…It is known that the performance of computational systems does not scale linearly as we increase the number of processing cores. We replaced the optimistic MultiExplorer approach for performance estimates by a SVM‐based regressor to CPU performance estimates systems 34 and a Decision Tree‐based regressor to GPUs 9,35 …”