Analyzing Traces from a Google Data Center

Minet, Pascale; Renault, Éric; Khoufi, Ines; Boumerdassi, Selma

doi:10.1109/iwcmc.2018.8450304

Cited by 25 publications

(12 citation statements)

References 6 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Google made publicly available a set of traces collected in one of its DCs over a period of 29 days [7], [2]. This DC data set has been extensively studied by researchers [8], [9], [10], [11], [12]. The analysis of this data set can help researchers i) to make valid assumptions, ii) design more accurate models of jobs, tasks and machines, and iii) propose more efficient job scheduling algorithms.…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Evaluating the Upper Bound of Energy Cost Saving by Proactive Data Center Management

Milocco

Minet

Renault

et al. 2020

IEEE Trans. Netw. Serv. Manage.

Self Cite

View full text Add to dashboard Cite

Data Centers (DCs) need to periodically configure their servers in order to meet user demands. Since appropriate proactive management to meet demands reduces the cost, either by improving Quality of Service (QoS) or saving energy, there is a great interest in studying different proactive strategies based on predictions of the energy used to serve CPU and memory requests. The amount of savings that can be achieved depends not only on the selected proactive strategy but also on userdemand statistics and the predictors used. Despite its importance, it is difficult to find theoretical studies that quantify the savings that can be made, due to the problem complexity. A proactive DC management strategy is presented together with its upper bound of energy cost savings obtained with respect to a purely reactive management. Using this method together with records of the recent past, it is possible to quantify the efficiency of different predictors. Both linear and nonlinear predictors are studied, using a Google data set collected over 29 days, to evaluate the benefits that can be obtained with these two predictors.

show abstract

Section: Related Workmentioning

confidence: 99%

“…Classification has been used for machine configuration in [12], where a machine configuration is defined as a vector whose first component denotes the CPU capacity and the second one the memory capacity. Classification has also been applied to jobs.…”

Section: Related Workmentioning

confidence: 99%

Evaluating the Upper Bound of Energy Cost Saving by Proactive Data Center Management

Milocco

Minet

Renault

et al. 2020

IEEE Trans. Netw. Serv. Manage.

Self Cite

View full text Add to dashboard Cite

show abstract

“…Several papers cited use a publicly available set of traces collected in one of Google's DCs over a period of 29 days [14], [16]. This DC data set has been extensively studied by researchers [15], [17], [18], [19], [20]. The analysis of this data set can help researchers i) to make valid assumptions, ii) design more accurate models of jobs, tasks and machines, and iii) propose more efficient job scheduling algorithms.…”

Section: Related Workmentioning

confidence: 99%

Proactive Data Center Management Using Predictive Approaches

et al. 2020

Self Cite

View full text Add to dashboard Cite

Data Center (DC) management aims at promptly serving user requests while minimizing the energy consumed. This is achieved by turning off unnecessary servers to save energy and adapting the number of servers that are on to the time-varying and heterogeneous user requests. A great change in the number of servers on leads to a considerable management effort, also called control effort in the literature, which should be reduced as much as possible. Since feedback control can improve the performance of computing systems and networks, we propose to use it to achieve this dynamic capacity provisioning of the DC. In order to design this feedback control, first, we developed a dynamic model of the DC. The purpose of this paper is to design a feedback control strategy based on the DC model, able to optimize i) the Quality of Service, ii) the energy consumed and iii) the management effort. A simple Reactive open-loop Control which provides an amount of energy equal to the amount requested in the previous time interval is considered as a benchmark for comparison. Second, two feedback controls based on the balance equations of the DC are studied, namely i) Reactive Feedback Control providing an amount of energy equal to that provided by the reactive open-loop control but adding the accumulated demand that has not yet been served, and ii) Model Predictive Control optimizing a constrained cost that weights the management effort and the prediction error. Reactive Control, Reactive Feedback Control and Model Predictive Control are compared in terms of energy consumed, energy error and management effort. Quantitative results of the comparative performance evaluation are given, based on a data set collected from a real DC.

show abstract

“…A data set collected in a Google data center over a period of 29 days and published in [2], [3], has been analyzed by many researchers [9]- [14] for both classification and prediction. In [9], K-Means is used to classify the jobs into three categories according to their duration: short, medium and long.…”

Section: State Of the Artmentioning

confidence: 99%

“…The authors show that the Bayesian method improves the ARMA prediction error by 5% at 6 hours, and 8% at 12 hours. In [14], the features of jobs submitted in the Google data center are studied.…”

Section: State Of the Artmentioning

confidence: 99%

Cost Reduction Bounds of Proactive Management Based on Request Prediction

Milocco

Minety

Renault

et al. 2019

2019 International Conference on High Performance Computing &Amp; Simulation (HPCS)

Self Cite

View full text Add to dashboard Cite

Data Centers (DCs) need to manage their servers periodically to meet user demand efficiently. Since the cost of the energy employed to serve the user demand is lower when DC settings (e.g. number of active servers) are done a priori (proactively), there is a great interest in studying different proactive strategies based on predictions of requests. The amount of savings in energy cost that can be achieved depends not only on the selected proactive strategy but also on the statistics of the demand and the predictors used. Despite its importance, due to the complexity of the problem it is difficult to find studies that quantify the savings that can be obtained. The main contribution of this paper is to propose a generic methodology to quantify the possible cost reduction using proactive management based on predictions. Thus, using this method together with past data it is possible to quantify the efficiency of different predictors as well as optimize proactive strategies. In this paper, the cost reduction is evaluated using both ARMA (Auto Regressive Moving Average) and LV (Last Value) predictors. We then apply this methodology to the Google dataset collected over a period of 29 days to evaluate the benefit that can be obtained with those two predictors in the considered DC.

show abstract

Analyzing Traces from a Google Data Center

Cited by 25 publications

References 6 publications

Evaluating the Upper Bound of Energy Cost Saving by Proactive Data Center Management

Evaluating the Upper Bound of Energy Cost Saving by Proactive Data Center Management

Proactive Data Center Management Using Predictive Approaches

Cost Reduction Bounds of Proactive Management Based on Request Prediction

Contact Info

Product

Resources

About