A topology-aware performance monitoring tool for shared resource management in multicore systems

Abstract

Nowadays, performance optimization involves careful data and task placement to deal with parallel application needs with respect to the underlying hardware topology. Monitoring the application behavior provides useful information that still needs to be matched with the actual placement, for instance to understand whether bottlenecks are caused by the sequential code itself or by shared resources in parallel programs. We propose an insightful monitoring tool based on two cornerstones of hardware performance counters monitoring and hardware locality modeling, respectively named PAPI and hwloc. It enables a dynamic visual analysis of parallel applications’ phases at runtime, revealing their possibly variable and heterogeneous behaviors and needs. A purpose designed application shows that the topology-aware visual representation of hardware counters can help figuring out shared resource bottlenecks and ease the task placement decision process in runtime systems.