Capturing the impact of external interference on HPC application performance

HPC applications are large software packages with high computation and storage requirements. To meet these requirements, the architectures of supercomputers are continuously evolving and their capabilities are continuously increasing. Present-day supercomputers have achieved petaflops of computational power by utilizing thousands to millions of compute cores, connected through specialized communication networks, and are equipped with petabytes of storage using a centralized I/O subsystem. While fulfilling the high resource demands of HPC applications, such a design also entails its own challenges. Applications running on these systems own the computation resources exclusively, but share the communication interconnect and the I/O subsystem with other concurrently running applications. Simultaneous access to these shared resources causes contention and inter-application interference, leading to degraded application performance. Inter-application interference is one of the sources of run-to-run variation. While other sources of variation, such as operating system jitter, have been investigated before, this doctoral thesis specifically focuses on inter-application interference and studies it from the perspective of an application. Variation in execution time not only causes uncertainty and affects user expectations (especially during performance analysis), but also causes suboptimal usage of HPC resources. Therefore, this thesis aims to evaluate inter-application interference, establish trends among applications under contention, and approximate the impact of external influences on the runtime of an application. To this end, this thesis first presents a method to correlate the performance of applications running side-by-side. The method divides the runtime of a system into globally synchronized, fine-grained time slices for which application performance data is recorded separately. The evaluation of the method demonstrates that correlating application performance data can identify inter-application interference. The thesis further uses the method to study I/O interference and shows that file access patterns are a significant factor in determining the interference potential of an application. This thesis also presents a technique to estimate the impact of external influences on an application run. The technique introduces the concept of intrinsic performance characteristics to cluster similar application execution segments. Anomalies in the cluster are the result of external interference. An evaluation with several benchmarks shows high accuracy in estimating the impact of interference from a single application run. The contributions of this thesis will help establish interference trends and devise interference mitigation techniques. Similarly, estimating the impact of external interference will restore user expectations and help performance analysts separate application performance from external influence