Simulation Models Verification for Resilient Communication on a Highly Adaptive Energy-Efficient Computer

The utility of simulations depends on the confidence in the simulation implementation and its results. This study discusses the verification of the communication models in the parallel trace-driven simulation framework HAEC-SIM. As simulation input, a parallel application is executed and recorded on an existing HPC system. The simulation focuses on modeling the transfer times of point-to-point messages within the application and the indirect effects resulting in an output trace of application events of the simulated execution on a target platform. Consequently, via verification the mes- sage transfer times obtained with HAEC-SIM are compared with those of an independent implementation of the commu- nication models. Both implementations consider the number of hops, the size and the target system parameters for each message. During verification the following factors are var- ied: application benchmarks, network topologies, mapping strategies, and resilient communication models. Verification yields an almost perfect agreement: the transfer times differ for a tiny percentage (0.00019 %) of messages by a negligible deviation of one picosecond, which is the finest granularity of the time data type. This result strengthens the confidence in a correct implementation of the communication models in simulation.