A Stochastic Composite Model to Understand the Impact of Rare, Colossal Interference in HPC Systems

Performance modeling of bulk-synchronous HPC applications is challenging in the presence of rare, colossal interference. This difficulty is attributed to the lack of available stochastic models that directly account for the rare, but large nature of interference at scale. The proposed approach in this paper provides a framework to correctly quantify and predict the impact of rare, colossal interference at scale on the HPC application's performance, as represented by the interval length. The efficacy of our framework is evaluated using synthetic HPC application data that depict rare and colossal interference.