| Description |
To address the need of understanding and optimizing the performance of complex applications and achieving sustained application performance across different architectures, we need performance models and tools that could quantify the theoretical performance and the resultant gap between theoretical and observed performance. This thesis proposes a benchmark-driven Roofline Model Toolkit to provide theoretical and achievable performance, and their resultant gap for multicore, manycore, and accelerated architectures. Roofline micro benchmarks are specialized to quantify the behavior of different architectural features. Compared to previous work on performance characterization, these micro benchmarks focus on capturing the performance of each level of the memory hierarchy, along with thread-level parallelism(TLP), instruction-level parallelism(ILP), and explicit Single Instruction, Multiple Data(SIMD) parallelism, measured in the context of the compilers and runtime environment on the target architecture. We also developed benchmarks to explore detailed memory subsystems behaviors and evaluate parallelization overhead. Beyond on-chip performance, we measure sustained Peripheral Component Interconnect Express(PCIe) throughput with four Graphics Processing Unit(GPU) memory managed mechanisms. By combining results from the architecture characterization with the Roofline Model based solely on architectural specification, this work offers insights for performance prediction of current and future architectures and their software systems. To that end, we instrument three applications and plot their resultant performance on the corresponding Roofline Model when run on a Blue Gene/Q architecture. |
