The International Conference for High Performance Computing, Networking, Storage and Analysis
Raexplore: Enabling Rapid, Automated Architecture Exploration for Full Applications.
Authors: Yao Zhang (Argonne National Laboratory), Prasanna Balaprakash (Argonne National Laboratory), Jiayuan Meng (Argonne National Laboratory), Vitali Morozov (Argonne National Laboratory), Scott Parker (Argonne National Laboratory), Kalyan Kumaran (Argonne National Laboratory)
Abstract: We present Raexplore, a performance modeling framework for architecture exploration. Raexplore enables rapid, automated, and systematic search of architecture design space by combining hardware counter-based performance characterization and analytical performance modeling. We demonstrate Raexplore for two recent manycore processors IBM Blue-Gene/Q compute chip and Intel Xeon Phi, targeting a set of scientific applications. Our framework is able to capture complex interactions between architectural components including instruction pipeline, cache, and memory, and to achieve a 3–22% error for same-architecture and cross-architecture performance predictions. Furthermore, we apply our framework to assess the two processors, and discover and evaluate a list of potential architectural scaling options for future processor designs.