Presentation
A Generalized Statistics-Based Model for Predicting Network-Induced Variability
Event Type
Workshop
W
Benchmarks
Performance
Scientific Computing
Simulation
TimeMonday, 18 November 20192pm - 2:30pm
Location603
DescriptionShared network topologies, such as dragonfly, subject applications to unavoidable inter-job interference arising from congestion on shared network links. Quantifying the impact of congestion is essential for effectively assessing and comparing the application runtimes. We use network performance counter-based metrics for this quantification. We claim and demonstrate that by using a local view of congestion captured through the counters monitored during a given application run, we can accurately determine the run conditions and thereby estimate the impact on the application's performance. We construct a predictive model that is trained using several applications with distinctive communication characteristics run under production system conditions with a 91% accuracy for predicting congestion effects.
Archive