Workload analysis for next-generation cores
Tipo de documentoProjecte Final de Màster Oficial
Condiciones de accesoAcceso restringido por acuerdo de confidencialidad
With Big Data and analytics applications becoming the focus of the decade for computer architects, there has risen a strong need to measure, compare and evaluate the systems that run these applications. The impact of such applications on the microarchitecture is an analysis in progress in many top research groups. Machine learning and data analytics kernels form an important component of such applications. NU-MineBench is a set of representative kernels and workloads that capture data mining code behavior. This thesis presents a detailed analysis of of this workload suite with insights about single thread performance, an evaluation of bottlenecks and opportunities for performance improvements. Identifying true bottlenecks in complex out-of-order processors is a challenging task and we look to the Top-down Microarchitecture Analysis Methodology (TMAM) for this. Upon identifying these bottlenecks, we propose design choices as well as workload optimizations in order to mitigate them.