download0 view756
twitter facebook

공공누리This item is licensed Korea Open Government License

dc.contributor.author
황은지
dc.contributor.author
최영리
dc.contributor.author
김선태
dc.contributor.author
김직수
dc.contributor.author
황순욱
dc.date.accessioned
2019-08-28T07:41:54Z
dc.date.available
2019-08-28T07:41:54Z
dc.date.issued
2016-09-15
dc.identifier.issn
1386-7857
dc.identifier.uri
https://repository.kisti.re.kr/handle/10580/14524
dc.description.abstract
Loosely coupled applications composed of a potentially very large number (from tens of thousands to even billions) of tasks are commonly used in high-throughput computing and many-task computing paradigms. To efficiently execute large-scale computations which can exceed the capability in a single type of computing resources within expected time, we should be able to effectively integrate resources from heterogeneous distributed computing (HDC) systems such as clusters, grids, and clouds. In this paper, we quantitatively analyze the performance of three different real scientific applications consisting of many tasks on top of HDC systems based on a partnership of distributed computing clusters, grids, and clouds to understand the application and resource characteristics, and show practical issues that normal scientific users can face during the course of leveraging these systems. Our experimental study shows that the performance of a loosely coupled application can be significantly affected by the characteristics of a HDC system, along with hardware specification of a node, and their impacts on the performance can vary widely depending on the resource usage pattern of each application. We then devise a preference-based scheduling algorithm that can reflect characteristics and resource usage patterns of various loosely coupled applications running on top of HDC systems from our experimental study. Our preference-based scheduling algorithm can allocate the resources from different HDC systems to loosely coupled applications based on the preferences of the applications for the HDC systems. We evaluate the overall system performance over various preference types, using trace-based simulations, which can be determined based on different factors such as CPU specifications and application throughputs. Our simulation results demonstrate the importance of understanding the application and resource characteristics on effective scheduling of loosely coupled applications on the HDC systems.
dc.language
eng
dc.relation.ispartofseries
Cluster Computing(The Journal of Networks, Software Tools and Applications)
dc.title
On the role of application and resource characterizations in heterogeneous distributed computing sys
dc.subject.keyword
Loosely coupled applications
dc.subject.keyword
High-throughput computing
dc.subject.keyword
Many-task computing
dc.subject.keyword
Performance evaluation
dc.subject.keyword
Heterogeneous distributed computing systems
dc.subject.keyword
Scheduling algorithm
Appears in Collections:
7. KISTI 연구성과 > 학술지 발표논문
Files in This Item:
There are no files associated with this item.

Browse