Dr. Paul J. Kuehn

Speaker Biography:

Paul J. Kuehn received the Dipl.-Ing. Dr.-Ing. and Dr.-Ing.habil degrees from the University of Stuttgart , Germany, in 1967, 1972 , and 1981. From 1977 to 1978 he was Member of the  Technical Staff at Bell Laboratories in Holmdel, N:J, USA. From 1978 to 1983 he was appointed Full Professor at the University of Siegen and from 1983 to 2009 Full Professor  and Director of the Institute of Communication Networks and Computer Engineering at the University of Suttgart, Germany. His main research fields are network architectures, communication engineering, and stochastic performance evaluation. He introduced the first Double Diploma Program between the U Stuttgart and French Grande Ecoles for Telecommunications . He was appointed Professor Associe at the Grande Ecole ENST, Paris, 1991, and Founding Dean of the Faculty of Information Engineering and Technology at the German University in Cairo.in 2002. He is an Honorary Senator of the University of  Mannheim,Germany, for his responsibility for the founding of the Faculty of Computer  Engineering in 1989. He is the Founder of the International MSc.-Program INFOTECH at the University of Stuttgart in 1999 , with close to 1000 graduations since then. From 1993 to 1998 he was the chairman of the German Center  Research Program “Mobile Communications” of the German Research Council DFG and from 1993 to 2002 he coordinated he  interdisciplinary Graduate College Program “Distributed Systems and Communication”. From 1991 to 2007 he was the Chairman of the  International Advisory Council (IAC) of the International Teletraffic Congress (ITC). He is member of the German National Academy Leopoldina, the Heidelberger Academy of Science, and the German Technical Academy acatech. He received the Dr.h.c. degree from the  Lund University of Technology, Sweden, and Dr.-Ing.E.h. degree from the Dresden Technical University in Germany for his  contributions in his research fields. He was appointed IEEE Fellow in 1989 and IEEE Life Fellow in 2009. He is also the recipient of the Christo Colombus Gold Medal of the Council of the City of Genova, Italy, for his contributions in the field of telecommunications.  The French Government honored his activities in International Science Exchange by the admission to the Chevalier Ordre Palmes  Academiques in 2002. In 2003 he was the recipient of the Eduard Rhein Price for Pioneering Work in Teletraffic Theory and Packet  Switching Technology, Germany. In 2009 he was appointed Honorary Member of the Electrotechnical Society VDE in Germany. In  2010 he received the Arne Jensen Lifetime Achievement Award from the International Teletraffic Congress ITC, and in 2018 he was  appointed Fellow of the German Information Technology Society ITG.

Talk Title:

Optimization of Energy Consumption of Cloud Data Center Server Clusters by Performance Modeling

Talk Abstract:

Data are the “raw material” of our information society which are stored in distributed Cloud Data Centers and can be accessed through the Internet for applications on business, production, logistic, and administrative control processes. The amount of data and the internet communication traffic grow exponentially, doubling roughly every 18 months and contributing by about 2% to the total global energy consumption and its corresponding carbon dioxide production. Optimizing this energy consumption can therefore contribute to a sustainable usage. There are several principal methods for the reduction of data center energy consumption as server consolidation, low-power operation or “sleep” modes, Dynamic Voltage and Frequency Scaling (DVFS), and Dynamic Load Balancing (DLB) through virtual machine migration. The corresponding control functions have to be additionally subjected to performance restrictions to meet application process Quality of Service (QoS) objectives. Typical approaches cover automatic DVFS through power consumption sensing and supply voltage adaptation, experimental tests through measurements, computer simulation, or mathematical parameter optimizations which provide performance results. These methods, however, are lacking with respect to run-time overhead or to requirements to meet prescribed QoS objectives. Our approach aims therefore on modeling and stochastic performance evaluations by which the cloud DC server clusters are modeled by queuing systems which provides a principally deeper insight on parametric dependencies. Three principally different approaches are studied:
Server Consolidation (SC) methods aim at sleep or low power consumption operation of idled servers. We model these configurations by a multi-dimensional system state description operating under an adaptive load-dependent hysteresis queue discipline which is optimized to meet prescribed Service Level Objectives (SLA) to guarantee either mean value or quantiles of service delays. The analysis is based on Markovian traffic arrival load and service assumptions; these results have been validated by computer simulations to be meaningful for more general traffic assumptions. The model is also applicable for low-power or sleep operation considering activation overhead of sleeping servers.
Dynamic Voltage and Frequency Scaling (DVFS) makes use of the required power consumption of electronic devices which depends linearly on clock frequency and device capacity, but increases quadratic with the supply voltage. For that reason the supply voltage is subdivided into a number of fixed operating ranges. DVFS is already automatically applied for server operation based on device heat parameters, which requires sensing and processing overhead, but cannot guarantee a prescribed SLA margin. We model these configurations by queuing systems from which we derive the exact arrival rate ranges of application jobs for each operating voltage level under prescribed SLA levels of QoS boundaries for either mean delays or delay quantiles. Applying our load-dependent hysteresis queue discipline for server consolidation, we can apply DVFS without any operating system interruption.
Dynamic Load Balancing (DLB) aims at the cooperation between cloud server clusters or whole cloud data centers by scheduling methods. This can be accomplished either through job migration between co-operating server clusters within a data center or by job migrations between different cloud data centers .Two different DLB strategies are modeled considering either the distance between cooperating data centers (Local Server System First (LSSF) or the response time (Shortest Response Time First (SRTF). Both models are analyzed by multi-dimensional Markovian state space descriptions including job migration overhead. Finally, a generally applicable method is applied which is based on the well-known effect of “Economy of Scale” by bundling server resources to a virtually aggregated server group.
For all modeling approaches the principal analysis methods are outlined with references to recent publications for details. For all methods sample results are presented, discussed, and how they can be applied for an optimized operation.

Keywords:

Cloud data centers, Energy efficiency, Server consolidation, Dynamic voltage and frequency scaling, Dynamic load balancing, Performance modeling, Queuing systems.