Shifting the Barrier: Extending the Boundaries of the BarrierPoint Methodology
Miguel Tairum Cruz, Sascha Bischoff, Roxana Rusitoru · 2018
The BarrierPoint methodology aims to reduce the system-level simulation cost of High-Performance Computing (HPC) systems by splitting workloads into small, representative sections. Whilst the BarrierPoint methodology is able to give realistic performance estimates for a wide range of applications, there are some cases in which it cannot be applied, such as workloads with too few barrier points, resulting in low run time improvements. In this work, we extend a previous cross-architectural BarrierPoint methodology, and demonstrate how to apply it in cases where it previously under-performed due to a small number of barrier points. We introduce variable barrier point granularity, enabling barrier point fine-tuning based on the different barrier point types. This leads to a further improvement of over 4.3x faster simulation times in addition to the previous BarrierPoint methodology, when comparing the same benchmarks.