Can’t See the Wood for the Trees? - Data Center Resource Management Based on Dimensionality Reduction of Workload DataZusammenfassung
Today’s data centers host thousands of business applications with varying workload behavior in virtual machines (VMs). The underlying infrastructures contain hundreds or even thousands of physical (host) servers. As energy usage - a major cost driver in data centers - increases with the server count, efficient capacity usage is key to operational eﬃciency. While observed workload traces (CPU-demand, memory-usage, etc.) would allow for accurate forecasting and optimal allocation of VM to physical servers, the sheer volume of data renders this task almost impossible for any but small server clusters. To cope with the managerial complexity, simple, conservative, and rather data-agnostic heuristics are applied. In this talk I will consider the general problem of analysing large volumes of workload data in order to extract signiﬁcant features (pattern) and use these features to allocate VMs eﬃciently to physical servers. By means of truncated Singular Value Decomposition an optimal orthonormal transform will be learned from the available workload data. The transform is then used to derive a compact representation of the original high-dimensional data. Application one will be a parsimonious, scalable optimization model to automatically determine resource allocation based on the reduced description of workload data. Application two will be a decision support tool that allows for a concise identification of periods of stable resource demands and turning points in VM workload behavior that require VM migrations.
Thomas Setzer received his Dipl.-Wi.-Ing. in Business Engineering from a research institute formerly known as University of Karlsruhe, and his Dr. rer. nat. in Information Systems from the Technische Universität München (TUM). He worked as a consultant and engineer for the European Parliament and Lufthansa. In 2007, Thomas was working as Visiting Scientist at IBM Research, New York. Currently, Thomas is heading the IT Service Operations Management Research Group at TUM. For his current work on data center virtualization and automation topics - in collaboration with Siemens IT Solutions and Services (SIS) and funded by the German Research Foundation (DFG) - he won the INFORMS ISS Design Science Award 2008.