Impact of Urgent Computing on Resource Management Policies, Schedules and Resources Utilization

详细信息	查看全文 | 推荐本文 |

• 作者：Krzysztof Kurowski^a ; ^{krzysztof.kurowski@man.poznan.pl} ; Ariel Oleksiak^a ; Wojciech Piatek^a ; Jan W臋glarz^a ; ^b
• 关键词：urgent computing ; on-demand computing ; parallel job scheduling ; hierarchical scheduling ; HPC ; Grids
• 刊名：Procedia Computer Science
• 出版年：2012
• 期刊代码：pc74_18770509
• 类别：cp
• 出版时间：2012
• 卷：9
• 期：Complete
• 页码：1713-1722
• 文件大小：1295 K

摘要

In general resource management is often defined as a process of identifying application requirements, matching compute resources to jobs, allocating those resources, scheduling and monitoring them over time to run jobs as effi-ciently as possible. In case of urgent computing, additionally, in this process we have to deal with a set of specific event-driven or time-critical application requirements that must be satisfied by a resource provider to guarantee the immediate access to compute resources. There is no question that capabilities supporting urgent computing are re-quired in distributed high performance and high throughput computing infrastructures like Grids. Nevertheless, it is difficult to predict if and how urgent jobs will disturb scheduling policies as well as coexist with less critical waiting and running jobs submitted by other end-users. To date many approaches have been proposed to deal with urgent com-puting, for instance based on preemption or high job priorities and recently based on advance reservations. However, there is a lack of comprehensive analysis of the impact of urgent computing on existing schedulers and their effi-ciency. Moreover, many resource providers use a single evaluation criterion -utilization or average load of compute resources. In fact the utilization of compute resources as a metric is in contrast to other evaluation criteria relevant for urgent jobs, e.g. waiting time, response time, or mean fiow time. These additional criteria in our opinion can be treated as good metrics for urgent computing taking into account the end-users perspective rather than the resource provider centric view. As we demonstrate in this paper, the problem is not trivial and it requires workload analyses. Finally, in this paper we successfully used GSSIM simulator to run various experiments and benchmarks for different scheduling strategies where urgent jobs were additionally considered.