TOWARDS EFFICIENT MAPPING IN LARGE-SCALE TRACE-DRIVEN PARALLEL PERFORMANCE SIMULATION

摘要

In parallel performance simulation of parallel systems, a large amount of Logic Processes (LP) must be mapped to relatively small number of Physical Elements (PE). Previous researches have shown that different mapping schemes could result in significant variation in the whole parallel simulation cost. In this paper, we propose, implement, and evaluate a Minimum Communication-guided Mapping (MiniCoM) scheme for large-scale trace-driven parallel performance simulation. Guided by extracted information about interactions among LPs from previously generated traces, MiniCoM can map some most frequently interacted LPs to the same PE while trying to keep load balance between PEs. The mapping aims to minimize realistic communications among PEs which may run on different nodes of host systems with large inter-node latency. We use BigSim simulator and two target programs to evaluate MiniCoM. Our results show that MiniCoM is more efficient than blocked mapping adopted by BigSim: it can reduce the total parallel simulation runtime by up to 49％.