The Effects of System Management Interrupts on Multithreaded, Hyper-threaded, and MPI Applications
Konstantin Macarenco, Kristina Frye, Benjamin Hamlin, Karen L. Karavanic · 2016
System Management Interrupts (SMIs) are a potential source of noise for parallel applications at all levels, and their use may be increasing. Building on a previous effort, we have conducted a study of SMI effects on parallel applications. In this paper we present initial results for NAS benchmarks, a multithreaded application, and the UnixBench benchmark. Findings reconfirm that performance degradation scales with the latency of the SMI. Additional results show that performance degradation increases when SMIs are enabled upon multiple communicating nodes, and may be amplified with HTT. This potential source of noise is of interest to tool developers both to detect its presence and also to accommodate the effect on measurement tools.