download0 view152
twitter facebook

공공누리This item is licensed Korea Open Government License

Title
Runtime prediction of parallel applications with workload-aware clustering
Author(s)
박주원김은혜
Publication Year
2017-11-01
Abstract
Several fields of science have demanded large-scale workflow support, which requires thousands of CPU cores or more. In order to support such large-scale scientific workflows, high capacity parallel systems such as supercomputers are widely used. In order to increase the utilization of these systems, most schedulers use backfilling policy: Small jobs are moved ahead to fill in holes in the schedule when large jobs do not delay. Since an estimate of the runtime is necessary for backfilling, most parallel systems use user's estimated runtime. However, it is found to be extremely inaccurate because users overestimate their jobs. Therefore, in this paper, we propose a novel system for the runtime prediction based on workload-aware clustering with the goal of improving prediction performance. The proposed scheme develops support vector regression model by the clusters resulted from a self-organizing map and hierarchical clustering analysis with the feature space reduced by factor analysis to reinforce prediction accuracy. In the experiments, we use workload logs on parallel systems (i.e., iPSC, LANL-CM5, SDSC-Par95, SDSC-Par96, and CTC-SP2) to evaluate the effectiveness of our approach. Comparing with other techniques, experimental results show that the proposed method improves the accuracy up to 69.08%.
Keyword
Runtime prediction; workload-aware clustering; support vector regression; machine learning approach
Journal Title
The Journal of Supercomputing
Citation Volume
73
ISSN
0920-8542
Files in This Item:
There are no files associated with this item.
Appears in Collections:
7. KISTI 연구성과 > 학술지 발표논문
URI
https://repository.kisti.re.kr/handle/10580/14612
http://www.ndsl.kr/ndsl/search/detail/article/articleSearchResultDetail.do?cn=NART78984346
Export
RIS (EndNote)
XLS (Excel)
XML

Browse