Improving Multi-threaded Application Performance on Modern Chip Multiprocessors

Multi-threading is one way of leveraging the increasing number of cores in today's processors to speed up applications by dividing the workload, that would otherwise be executed sequentially, across multiple threads that run in parallel. However, if one thread lags behind the rest of the threads, it will limit the overall achievable benefit from parallelization. To unlock the full potential of multi-threading such critical threads have to be predicted and optimized during runtime in a way that minimizes the difference in performance between them and the remaining threads. The large number of factors that can affect a thread's performance make critical thread prediction a very challenging task. Moreover, these factors are not completely independent of each other and can be indirectly related in ways that are difficult to model accurately. Therefore, previous works that addressed the issue of thread criticality only focused on a limited set of factors when predicting which threads are critical. However, the results presented in this thesis show that none of these factors strongly correlates to thread criticality when considered separately resulting in suboptimal thread criticality predictions. Instead of focusing on a limited set of factors, this thesis leverages machine learning to facilitate the study of a wide variety of factors that can potentially affect a thread's progress. Specifically, this thesis formulates the problem of thread criticality as a Learning-to-Rank problem. Each thread is represented by a set of runtime statistics representing the input features to the ranking model and the predicted scores reflect how critical a thread is. Using a ranking model puts more emphasis on threads' criticality scores relative to each other rather than exactly matching the criticality score for each thread. This work shows how the proposed thread criticality prediction methodology can be applied to two types of run environments. The first runs the benchmarks on a real system and uses Linux Trace Toolkit: next generation (LTTng) for data collection while the second uses Gem5 to simulate the required multi-core system. Collected results show that single features are not very accurate predictors for thread criticality compared to a model that considers several features. The obtained model performs 17% better than a single-metric based predictor in predicting threads' criticality. Finally, a case study where critical threads are prioritized by doubling the corresponding core frequency shows the potential benefit the proposed ranking model offers compared to a predictor that is based on a single feature.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.