Performance Analysis of One-sided Communication Mechanisms
Bernd W. Mohr, Andrej Kühnal, Marc-André Hermanns, Felix Wolf · JuSER (Forschungszentrum Jülich) · 2005
Performance analysis of parallel programs requires information about the dynamic behavior of all participating processes.The dynamic behavior can be modeled as a stream or trace of events.The events are chosen in such a way that they represent important aspects in the execution of the application on a level of abstraction suitable for the analysis task.Based on this idea, the KOJAK toolkit for performance analysis records and analyzes the activities of MPI-1 point-to-point and collective communication.This paper describes the integration of performance measurement and analysis methods for remote memory access (RMA) or one-sided communication into the KOJAK toolkit, in particular for the MPI-2 and SHMEM interfaces.We introduce the underlying event model used to represent the dynamic behavior of RMA operations and show that our model reflects the relationships between communication and synchronization more accurately than existing models.Then, we present event patterns which are used by KOJAK to locate inefficient situations in a program's dynamic remote memory access behavior.