x

Data centers are the core of modern business environments as computation has been rapidly moving into the cloud in the last decade. Data centers are among the fastest growing users of electricity in the U.S. consuming an estimated 91 billion kilowatt-hours of electricity in 2013, about 2% of the total electricity consumed by the US. When operating a data center of hundreds of thousands of servers, it is essential that they be operated effectively to improve energy efficiency and environmental sustainability. Due to aggressive adoption of cloud-based computing, the demands on data centers are growing exponentially, and both academia and industry will need to rethink how data centers are designed, built, and operated to be sustainable. Despite a decade of research and industrial innovation, a recent report from Natural Resources Defense Council (NRDC) indicates typical small and mid-size data centers hosting private clouds still hold many wasteful practices. Whereas best practices at mega-scale commercial cloud operators (e.g., Facebook, Microsoft, Google, and Amazon) have addressed the most egregious wastes (e.g., inefficient cooling), we nevertheless must find ways to transfer these best practices across the data center landscape and address the remaining performance and efficiency challenges that afflict even the largest installations.

At NSF, several core and crosscutting programs, including CSR, CyberSEES and CPS, have taken action and invested in numerous projects nationwide to address the fundamental issues of sustainable data centers (SDC). These projects can be broadly classified into three areas: individual servers, data center-level resource management, and energy supply. For individual servers, the ultimate goal is to design energy-proportional computing nodes by reducing both idle power (e.g., PowerNap) and dynamic power (e.g., DVFS). For data center-level resource management, the main goal is to utilize the resources (i.e., CPU, memory, bandwidth) in an effective and efficient way, without wasting resources. A better understanding of the workload is the key. Cloud providers can allocate right-sized server and network platforms to meet users’ application requirements. At the energy supply, green power, such as solar and wind, are beginning to enter the data center power supply chain. By leveraging renewable power sources, local micro-grids may offset some or all of a data center's energy needs, particularly for small and mid-scale facilities. At the largest scale, incentives and mechanisms must be sought to encourage provisioning clean, reliable power in concert with the existing public grid.

Around the mid-2000’s, the advent of mega-scale internet services and public cloud offerings led to a redesign of data center architectures, which addressed key inefficiencies, particularly in electrical and mechanical infrastructure. At the same time, accelerated need for efficient servers spurred a generation of research on CPU, memory, network, and storage power management techniques, which have led to a marked improvement in server efficiency and energy proportionality. However, this first generation of improvement has plateaued; further opportunity in the large-scale mechanical infrastructure is limited and no single server or network component stands out as the key source of inefficiency. Hence, it is time for a second, holistic, clean-slate redesign of the data center, encompassing new server architectures, heterogeneous computing platforms, radical networking paradigms, new mechanical and electrical designs, intelligent cluster management, and radical rethinking of software architectures while considering changing usage patterns (e.g., hybrid private/public clouds).

Although the need for broad input on sustainable data center design is acute, concerns about competitive advantage and user privacy have made open collaboration between academic researchers and cloud operators difficult. Academic researchers have limited access to production data center facilities and hence are not always aware of the real problems faced by practitioners. The immediate risk of this disconnect is that researchers might spend their time attacking imagined problems that are irrelevant to modern practice. One of the goals of the proposed workshop is to rectify this disconnect by providing an open forum for researchers and industry practitioners to come together. In addition, a key objective of the workshop is to identify the opportunities, needs, and requirements to leverage NSF’s investment in the two NSFCloud testbeds to better serve energy-related research activities on data centers.

In addition to developing promising technologies to improve data center efficiency, we also need new metrics to assess the success of SDC research. Currently, power usage effectiveness (PUE) is a widely reported metric to assess the energy efficiency of a data center. The impact of renewables can be assessed via carbon usage effectiveness (CUE) to measure the combined impact of clean energy and energy efficiency on greenhouse gas emissions, and water usage effectiveness (WUE) can be used to assess the water usage of a data center. And yet, all three of these metrics fall short of describing the true efficiency of the data center. They fail to reflect waste at the enclosure/tray level (e.g., VRMs, server fans). Moreover, they do not assess the efficiency or value of the computation being performed and hence fail to reflect server hardware inefficiencies or software bloat.

In this workshop, we intend to bring together industry practitioners and academic researchers to build the community and discuss the vision, challenges, and opportunities for SDC research for the next 5-10 years. More specifically, the objectives of the workshop include: