Citation Context

...ies of modeling clusters make this impractical. Utilization-based large-scale simulation: Given these challenges, in this paper, we use a trace-driven simulation approach for data center environments =-=[27]-=-[29]. This approach uses real-world traces from actual enterprise deployments to drive individual server simulations. High-level models like those in Figure 5 are used to correlate resource utilizatio...

Citation Context

...guration, usage and even the battery of a mobile system and therefore motivate the proposed self-modeling approach. Out of simplicity, we use the 1 Hz energy model based on CPU utilization as used in =-=[18, 24]-=-. Our target system is a Lenovo ThinkPad T61 laptop. We measure its power with a USB2533 Data Acquisition System from Measurement Computing along with another PC. 3.1 Dependency on Hardware Energy mod...

Citation Context

...the servers under test. The power consumption of the servers and the application throughput were collected for all the operation frequencies. More detailed description on the approach is available in =-=[4]-=-. The data from the experiments suggest that for a fixed operating frequency, both the power consumption of the server and the application throughput are approximately linear functions of the utilizat...

Citation Context

...erformance and power and a methodology for porting a model generated on one machine to another in the same family by examining the different relationships between performance and resource utilization =-=[55]-=-. Heath et al. used a similar approach to full-system power modeling in order to enable energy-aware server consolidation in clusters [27]. Using OS-reported CPU, memory, and disk utilization, they de...

Citation Context

...system power model is proposed in [33] that uses only processor level event counters, the hypothesis being that memory and I/O events are correlated with certain classes of processor level events. In =-=[25]-=-, both power and performance of an application are profiled to build a lookup table which is used to predict power and performance online. A common limitation of all these works is that they assume an...

Citation Context

...erformance and power and a methodology for porting a model generated on one machine to another in the same family by examining the different relationships between performance and resource utilization =-=[77]-=-. Heath et al. [34] used a similar approach to full-system power modeling to enable energy-aware server consolidation in clusters. Using OS-reported CPU, memory, and disk utilization, they developed l...

Citation Context

...ies of energy models on the configuration, usage and even the battery of a mobile system and therefore motivate the proposed selfmodeling approach. Out of simplicity, we use the 1Hz energy model from =-=[20, 24]-=- based on CPU utilization. Our target system is a Thinkpad T61 laptop. We measure its 2power with a USB-2533 Data Acquisition System from Measurement Computing along with another PC. 3.1 Dependency o...

Citation Context

...ly depends on the training data. In this work, we measured both of frequency and CPU-power using precise tools. There are several works considering power modeling of a full-server. Ranganathan et al. =-=[24]-=- has proposed a dynamic power budgeting optimizations. In their work, a lookup table has been built for relating power and performance to system resource utilization. Similarly, Fan et al. [10] have d...

Citation Context

...ng power models in related work. Section 3 describes our hardware and software experimental infrastructure. Section 4 presents the model variability both in feature selection and cluster power model prediction. We conclude with Section 5. 2 Related Work Previous studies model the power consumption of single nodes using different predictors and modeling techniques [3,4,6,12,14,16-18,20-23]. Some studies predict power consumption based on CPU utilization alone [6,18], while others use board-level measurements [16]. The modeling techniques also vary in complexity, from simple lookup-based models [20] to chaotic attraction predictors [16]. However, all of these studies build and validate models on a single node, assuming that these models can then be applied to other identically configured nodes without requiring refitting. We challenge that assumption in this work. Other studies use different validation techniques. Li and John validate their routine-specific models on a full-system simulator [17], which again assumes no inter-node variability. Vasan et al. present power measurements from a medium-scale datacenter but only build single-node models [25]. Heath et al. model the total power o...