Crossing the architectural barrier: Evaluating representative regions of parallel HPC applications

摘要

Exascale computing will get mankind closer to solving important social, scientific and engineering problems. Due to high prototyping costs, High Performance Computing (HPC) system architects make use of simulation models for design space exploration and hardware-software co-design. However, as HPC systems reach exascale proportions, the cost of simulation increases, since simulators themselves are largely single-threaded. Tools for selecting representative parts of parallel applications to reduce running costs are widespread, e.g., BarrierPoint achieves this by analysing, in simulation, abstract characteristics such as basic blocks and reuse distances. However, architectures new to HPC have a limited set of tools available. In this work, we provide an independent cross-architectural evaluation on real hardware - across Intel and ARM - of the BarrierPoint methodology, when applied to parallel HPC proxy applications. We present both cases: when the methodology can be applied and when it cannot. In the former case, results show that we can predict the performance of full application execution by running shorter representative sections. In the latter case, we dive into the underlying issues and suggest improvements. We demonstrate a total simulation time reduction of up to 178x, whilst keeping the error below 2.3% for both cycles and instructions.