CSS: Container Resource Manager Using System Call Pattern for Scientific Workflow

Multiple containers running scientific workflows in SMP-based high-performance computers generate some bottlenecks due to workload flexibility. To improve system resource utilization by minimizing these bottlenecks, vertical resource management is required to determine an appropriate resource usage...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Applied sciences 2022-08, Vol.12 (16), p.8228
Hauptverfasser: Song, Chunggeon, Yu, Heonchang, Lee, Eunyoung
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Multiple containers running scientific workflows in SMP-based high-performance computers generate some bottlenecks due to workload flexibility. To improve system resource utilization by minimizing these bottlenecks, vertical resource management is required to determine an appropriate resource usage policy according to the resource usage type of the container. However, the traditional methods have additional overhead for collecting monitoring metrics, and the structure of the resource manager is complex. In this paper, in order to compensate for these shortcomings, we propose CSS, a dynamic resource manager utilizing system call data collected for security purposes. The CSS utilizes the SBCC algorithm, which uses the number of futex system calls as a heuristic measure to determine the number of IO-intensive workload occurrences. In addition, the CTBRA algorithm is used to determine the range of resources to be allocated for each container and to perform actual resource allocation. We implemented a prototype of CSS and conducted experiments on NPB to analyze the performance of CSS with various types of large-scale tasks of a scientific workflow. As a result of the experiment, it showed a performance improvement of up to 7% compared with the environment where Linux cgroups were not applied. In addition, CANU performance analysis was performed to verify the effectiveness of applications used in the real world, and performance improvement of up to 4.5% was shown.
ISSN:2076-3417
2076-3417
DOI:10.3390/app12168228