Stochastic Model Predictive Control for Data Centers
Independent thesis Advanced level (professional degree), 20 credits / 30 HE creditsStudent thesis
Datacenters operations are notoriously energy-hungry with the computing and cooling infrastructures drawing comparable amount of power. A direction to improve their efficiency is to jointly control the Information Technology (IT) and Cooling Techniques (CT) components so that less cooling power has to be spent for the same Quality of Service (QoS) level. This work investigates minimum cost control strategies for this joint control problem at the computer room level by controlling the room’s cooling units and at the server level by controlling the server’s fans. The state of the art is too conservative and leads to over provision of mass flows and of the cooling power and this, in turn, leads to low efficiency; Moreover the existing literature does not include uncertainties in the models. We present a novel Stochastic Model Predictive Control (SMPC) strategy that sets the optimal mass-flow of each individual fan based on some predicted IT loads, the latter computed through a dedicated stochastic linear forecaster capturing the complex seasonal patterns that usually are exhibited by these quantities. The proposed controller thus manipulates the operation of the fans while simultaneously considering uncertainties in the forecasted IT loads and accounts for the different time-scales that characterize the thermal dynamics. Our simulations results show how it is possible to reduce the energyconsumption with respect to the existing practices at the same QoS levels while being robust to sudden changes in the IT loads.
Place, publisher, year, edition, pages
2016. , 63 p.
Technology, SMPC, MPC, Data center
IdentifiersURN: urn:nbn:se:ltu:diva-45847Local ID: 38192b42-429d-4590-a44b-c98a636a8d8bOAI: oai:DiVA.org:ltu-45847DiVA: diva2:1019145
Subject / course
Student thesis, at least 30 credits
Computer Science and Engineering, master's level
Validerat; 20160905 (global_studentproject_submitter)2016-10-042016-10-04Bibliographically approved