Datum: 30. Januar 2019
KIT, Fakultät für Informatik
Erscheinungsort / Ort: Karlsruhe
Referent(en): Hartmut Schmeck, Wolfgang Karl
Modern computing systems are becoming increasingly complex. On one end of the spectrum, personal computers now commonly support multiple processing cores, and, on the other end, Internet services routinely employ thousands of servers in distributed locations to provide the desired service to its users. In such complex systems, concerns about energy usage and power consumption are increasingly important. Moreover, growing awareness of environmental issues has added to the overall complexity by introducing new variables to the problem. In this regard, the ability to abstractly focus on the relevant details allows model-based design to help significantly in the analysis and solution of such problems. In this dissertation, we explore and analyze model-based design for energy and power considerations in computing systems. Although the presented techniques are more generally applicable, we focus their application on large-scale Internet services operating in U.S. electricity markets. Internet services are becoming increasingly popular in the ICT ecosystem of today. The physical infrastructure to support such services is commonly based on a group of cooperative data centers (DCs) operating in tandem. These DCs are geographically distributed to provide security and timing guarantees for their customers. To provide services to millions of customers, DCs employ hundreds of thousands of servers. These servers consume a large amount of energy that is traditionally produced by burning coal and employing other environmentally hazardous methods, such as nuclear and gas power generation plants. This large energy consumption results in significant and fast-growing financial and environmental costs. Consequently, for protection of local and global environments, governing bodies around the globe have begun to introduce legislation to encourage energy consumers, especially corporate entities, to increase the share of renewable energy (green energy) in their total energy consumption. However, in U.S. electricity markets, green energy is usually more expensive than energy generated from traditional sources like coal or petroleum. We model the overall problem in three sub-areas and explore different approaches aimed at reducing the environmental foot print and operating costs of multi-site Internet services, while honoring the Quality of Service (QoS) constraints as contracted in service level agreements (SLAs). First, we model the load distribution among member DCs of a multi-site Internet service. The use of green energy is optimized considering different factors such as (a) geographically and temporally variable electricity prices, (b) the multitude of available energy sources to choose from at each DC, (c) the necessity to support more than one SLA, and, (d) the requirements to offer more than one service at each DC. Various approaches are presented for solving this problem and extensive simulations using Google’s setup in North America are used to evaluate the presented approaches. Second, we explore the area of shaving the peaks in the energy demand of large electricity consumers, such as DCs by using a battery-based energy storage system. Electrical demand of DCs is typically peaky based on the usage cycle of their customers. Resultant peaks in the electrical demand require development and maintenance of a costlier energy delivery mechanism, and are often met using expensive gas or diesel generators which often have a higher environmental impact. To shave the peak power demand, a battery can be used which is charged during low load and is discharged during the peak loads. Since the batteries are costly, we present a scheme to estimate the size of battery required for any variable electrical load. The electrical load is modeled using the concept of arrival curves from Network Calculus. Our analysis mechanism can help determine the appropriate battery size for a given load arrival curve to reduce the peak. Third, we present techniques to employ intra-DC scheduling to regulate the peak power usage of each DC. The model we develop is equally applicable to an individual server with multi-/many-core chips as well as a complete DC with an intermix of homogeneous and heterogeneous servers. We evaluate these approaches on single-core and multi-core chip processors and present the results. Overall, our work demonstrates the value of model-based design for intelligent load distribution across DCs, storage integration, and per DC optimizations for efficient energy management to reduce operating costs and environmental footprint for multi-site Internet services.
DOI Link: 10.5445/IR/1000091774