Section: New Results

Machine Learning for an efficient and dynamic management of data centers

Data Analysis in Data Centers

Participants : Eric Renault ( Telecom Sud-Paris), Selma Boumerdassi ( Cnam), Pascale Minet, Ines Khoufi.

In High Performance Computing (HPC), it is assumed that all machines are homogeneous in terms of CPU and memory capacities, and that the tasks making up the jobs have similar resource requests. It has been shown that this homogeneity relating both to machine capacity and workload, although generally valid for HPC, does no longer apply to data centers. This explains why the publication of data gathered in an operational Google data center over 29 days has aroused great interest among researchers.

It is crucial to have real traces of a Google data center publicly available that are representative of the functioning of real data centers. Our goal is to analyze the data collected and to draw useful conclusions about machines, jobs and tasks as well as resource usage. Our main results have been published in  [25], [24] and can be summarized as follows:

• Although $92\%$ of machines have a CPU capacity of $0.5$, there are 10 machine configurations in the data center, each configuration is characterized by a pair $(CPUcapacity,memorycapacity)$. The most frequent configuration is supported by only $53\%$ of machines.

• Over the 29 days, all the machines in the data center that were removed, were restarted later after an off-period. $50\%$ of these periods have a duration less than or equal to 1000 seconds (i.e. $16.66$ minutes), suggesting a maintenance operation.

• The distribution of jobs per category reveals only one job, representing $0.002\%$, for the Infrastructure, $0.13\%$ of jobs for Monitoring, $9.91\%$ of jobs for Production, $56.30\%$ of jobs for Other, and $33.63\%$ of jobs for Free. $92.05\%$ of jobs have a single task. $95.75\%$ have fewer than 10 tasks. But 12 jobs have 5000 tasks and 114 jobs have around 1000 tasks.

• With regard to resource requests, $0.11\%$ of jobs have a memory request and a CPU request higher than or equal to $10\%$.

• $94.25\%$ of jobs wait less than 10 seconds before being scheduled. However, some of them wait for more than 1000 seconds. Such large values could be explained by the existence of placement constraints for the jobs, making them harder to place and schedule. $49\%$ of jobs have an execution time less than 100 seconds.

Such results are needed to validate or invalidate some simplifying assumptions that are usually made when reasoning about models, and make the models more accurate for jobs and tasks as well as for available machines. Having validated these models on real data centers, they can then be used for extensive evaluation of placement and scheduling algorithms and more generally for resource allocation (i.e. CPU and memory). These algorithms can then be applied in real data centers.

Another possible use of this data set is to consider it as a learning set in order to predict some feature of the data center, such as the workload of hosts or the next arrival of jobs.

Machine Learning for an Energy-Efficient Management of Data Centers

Participants : Ruben Milocco ( University Of Camahue, Argentina), 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. The problem of managing Data Centers (DC) and clouds optimally, in the sense that the demand is met with a minimal energy cost, remains a major issue. In this research, we evaluate the maximum energy saving that can be obtained in DCs by means of a proactive management of resources. The proposed management is based on models that predict resource requests.

Diverse approaches to obtain predictive models of DCs have been studied recently. Among the most popular methods with the comparatively lowest prediction errors are the predictive models of the ARMAX family. Hence, we study the predictive model given by the ARMAX family. We compare its performance with that of the Last Value (LV) model which predicts that the next value will be equal to the current one. To the best of our knowledge, there are no studies relating to the performance bounds that can be achieved using these models. In this research, we study the limits of the improvement in terms of energy cost that can be obtained using proactive strategies for DC management based on predictive models.

Using the Google dataset collected over a period of 29 days and made publicly available, we evaluate the largest benefit that can be obtained with those two predictors.