StatTask: Reuse distance analysis for task-based applications
2015 (English)In: Proc. 7th Workshop on Rapid Simulation and Performance Evaluation: Methods and Tools, New York: ACM Press, 2015, 1-7 p.Conference paper (Refereed)
Task-based programming has grown in popularity as it provides programmers with an intuitive abstraction for expressing parallelism, and runtimes with flexibility in scheduling and load-balancing. However, while good tools exist for understanding scheduling and load-balancing, little work has been done to analyze how tasks and schedules interact with the cache hierarchy. This area of investigation is particularly important as achieving good data reuse through complex cache hierarchies is essential for performance on modern systems, and the inherent flexibility of task-based runtimes offers an exciting potential for applying such knowledge to improve scheduling.
This work presents a new approach to model the interaction between tasks, schedules, and the memory hierarchy. The model analyzes the tasks’ data reuse to predict cache behavior for arbitrary task schedules on different memory hierarchies, and does so from only one profiling run. With this approach we can identify the potential for data reuse between tasks, and which can then be leveraged to improve scheduling.
Place, publisher, year, edition, pages
New York: ACM Press, 2015. 1-7 p.
IdentifiersURN: urn:nbn:se:uu:diva-242548DOI: 10.1145/2693433.2693434ISBN: 978-1-60558-699-1OAI: oai:DiVA.org:uu-242548DiVA: diva2:783829
RAPIDO 2015, January 21, Amsterdam, The Netherlands
ProjectsResource Sharing Modeling
FunderSwedish Research Council