Citation Context

...haracterization of a web server by Iyenger et al. [19] and workload models of a large scale server farm by Bent et al [3] fall in the first category. Individual server utilization has been studied in =-=[5, 16, 4]-=-. In [5], Bohrer et al use peak-trough analysis of commercial web servers to establish that the average CPU utilization for typical web servers is fairly low. Similar observations on the peaktrough na...

Citation Context

... better support this scenario by minimizing migrations during successive control intervals. Though enterprise application workloads often have time varying loads that behave according to patterns [23]=-=[8]-=-, actual demands are statistical in nature and are likely to differ from predictions. Therefore, to further improve the efficiency and application quality of service of our approach, we manage workloa...

Citation Context

...a shared hosting platform often hosts a moderate number of well-known services over long periods of time. It is also possible to build analytical models of resource needs and of their temporal trends =-=[60, 23, 49, 17]-=-, and even to augment these models to account for virtualization overheads [67]. Finally, another approach consists in monitoring services as they run to determine their needs [70, 53, 45, 32, 9]. [25...

Citation Context

... analytic methods, our approach has involved a full prototype implementation and experiments on an actual Linux cluster. k-means clustering is a common technique used for static workload analysis. In =-=[7]-=-, k-means is used for workload analysis and demand prediction. In constrast, we use k-means to automatically characterize the workload and use queueing theory approaches for provisioning. Further, our...

Citation Context

...identifies a few workload classes and reuses previous resource allocations for each class to minimize reallocation overheads [60]. CloudScale [56], PRESS [29], AGILE [46] and the work by Gmach et al. =-=[28]-=- perform online prediction of resource needs, often without a priori workload knowledge. Finally, auto-scaling systems such as Rightscale [52] automatically scale the number of physical or virtual ins...

Citation Context

...er. Recent research has focused on improving both SLA management and data center energy efficiency, mostly with respect to resource allocation [6], [7], capacity planning and workload management [8], =-=[9]-=-, power [10], [11] and thermal management [5], [12]. However, most published research focuses on managing performance, IT resources, workload, power and cooling in isolated ways. Approaches that are n...

Citation Context

...server will rise gradually from 0 to Rj . C. WORKLOAD PREDICTION Enterprise datacenters workloads typically show a repetitive pattern with a period in the order of hours, days, weeks and so forth. In =-=[14]-=- authors have demonstrated that for the purpose of workload forecasting, the period of workload behavior is equal to 7 days for a large variety of datacenter applications. The forecasting method we us...

Citation Context

...llenge that requires both sophisticated online algorithms and excellent predictions of future workloads. 1 1. Note that short-term predictions of workload demand within 24 hours can be quite accurate =-=[18]-=-, [23].9 % cost reduction 40 30 20 10 0 T=1s Poisson �=1.2 �=1.5 �=1.8 sec min hour day sec min hour day � / e 0 � / e 0 (a) Hotmail (b) MSR Fig. 5. Impact of burstiness on the cost savings of dynami...

Citation Context

...oad per generated task. This workload prediction ability is a key enabler for a datacenter resource manager. There are many techniques to do datacenter workload modeling and prediction. See reference =-=[57]-=- for a representative work on characterizing datacenter workload demand patterns. See reference [58] for a qualitative comparison of work on datacenter workload modeling based 9To appear in IEEE Tran...