Section: New Results

Machine Learning applied to Networking

Machine Learning for energy-efficient and QoS-aware Data Centers

Participants : Ruben Milocco ( Comahue University, Argentina, Invited Professor ), Pascale Minet, Eric Renault ( Telecom Sud-Paris ), Selma Boumerdassi ( Cnam ).

To limit global warming, all industrial sectors must make effort to reduce their carbon footprint. Information and Communication Technologies (ICTs) alone generate 2% of global CO2 emissions every year. Due to the rapid growth in Internet services, data centers have the largest carbon footprint of all ICTs. According to ARCEP (the French telecommunications regulator), Internet data traffic multiplied by 4.5 between 2011 and 2016. In order to support such a growth and maintain this traffic, data centers'energy consumption needs to be optimized.

We determine whether resource allocation in DCs can satisfy the three following requirements: 1) meet user requirements (e.g. short response times), 2) keep the data center efficient, and 3) reduce the carbon footprint.

An efficient way to reduce the energy consumption in a DC is to turn off servers that are not used for a minimum duration. The high dynamicity of the jobs submitted to the DC requires periodically adjusting the number of active servers to meet job requests. This is called Dynamic Capacity Provisioning. This provisioning can be based on prediction. In such a case, a proactive management of the DC is performed. The goal of this study is to provide a methodology to evaluate the energy cost reduction brought by proactive management, while keeping a high level of user satisfaction.

The state-of-the art shows that appropriate proactive management improves the cost, either by improving QoS or saving energy. As a consequence, there is great interest in studying different proactive strategies based on predictions of either the energy or the resources needed to serve CPU and memory requests. The cost depends on 1) the proactive strategy used, 2) the workload requested by jobs and 3) the prediction used. The problem complexity explains why, despite its importance, the maximum cost savings have not been evaluated in theoretical studies.

We propose a method to compute the upper bound of the relative cost savings obtained by proactive management compared to a purely reactive management based on the Last Value. With this method, it becomes possible to quantitatively compare the efficiency of two predictors.

We also show how to apply this method to a real DC and how to select the value of the DC parameters to get the maximum cost savings. Two types of predictors are studied: linear predictors, represented by the ARMA model, and nonlinear predictors obtained by maximizing the conditional probability of the next sample, given the past. They are both applied to the publicly available Google dataset collected over a period of 29 days. We evaluate the largest benefit that can be obtained with those two predictors. Some of these results have been presented at HPCS 2019 [20].

Machine Learning applied to IoT networks

Participants : Miguel Landry Foko Sindjoung ( Phd Student, Dschang University, Cameroon, Inria Internship), Pascale Minet.

Knowledge of link quality in IoT networks allows a more accurate selection of wireless links to build the routes used for data gathering. The number of re-transmissions is decreased, leading to shorter end-to-end latency, better end-to-end reliability and a longer network lifetime.

We propose to predict link quality by means of machine learning techniques applied on two metrics: the Received Signal Strength Indicator (RSSI) and the Packet Delivery Ratio (PDR). These two metrics were selected because RSSI is a hardware metric that is easily obtained and PDR takes into account packets that are not successfully received, unlike RSSI.

The data set used in this study was collected from a TSCH network deployed in the Grenoble testbed consisting of 50 nodes operating on 16 channels. Data collected by Mercator include 108659 measurements of PDR and average RSSI. We train the model over the training set and predict the link quality on the channel considered for the samples in the validation set. By comparing the predicted values with the real values, the confusion matrix is computed by evaluating the number of true-positive, true-negative, false-positive and false-negative for the link and channel considered.

Whatever the link quality estimator used, RSSI, PDR or both, the Random Forest (RF) classifier model outperforms the other models studied: Linear Regression, Linear Support Vector Machine, Support Vector Machine.

Since using Bad links that have been predicted Good strongly penalizes network performance in terms of end-to-end latency, end-to-end reliability and network lifetime, the joint use of PDR and RSSI improves the accuracy of link quality prediction. Hence, we recommend using the Random Forest classifier applied on both PDR and RSSI metrics. This work has been presented at the PEMWN 2019 conference [33].