Fast Modeling of Network Contention in Batch Point-to-point Communications by Packet-level Simulation with Dynamic Time-stepping
Yang Zhang, Jintao Peng, Qingkai Liu · 2020
Network contention has long been one of the root causes of performance loss in large-scale parallel applications. With the increasing importance of performance modeling to both large-scale application optimization and application-system co-design, the conflict of speed and accuracy in contention modeling is becoming prominent. Cycle-accurate network simulators are often too slow for large scale applications, while point-to-point analytical models are not accurate enough to capture the contention effects. To model the network contention in batch point-to-point communications, we propose a unified contention model after the flow-fair end-to-end congestion control mechanism. The model uses packet-level simulations to be accurate, but can be approximated by a flow-level semi-analytical model when messages are large enough, thus is fast. Furthermore, we propose a dynamic time-stepping technique which significantly speeds up the packet-level simulation with only minor accuracy loss. Experiments with typical communication patterns and application traces show that our model accurately predicates the communication time with an average error of 9%(fixed time step) and the dynamic time-stepping technique improve the simulation performance by up to 131 folds with an average accuracy loss of 10.5% for real application traces.