Large-scale computing can be used to characterize the host infrastructures of internet-services, social networks and cloud services, the growing pervasiveness of which has been undeniable in recent years. Although the utilization of these services has triggered significant growth in several areas such as science, communication, education and health-care, to name a few, it has not come without a cost. Specifically, the operation of large-scale infrastructures can result in energy consumptions equivalent to the energy budget of a medium-sized city. The deployment of such infrastructures requires hunders of kgs of primary materials and fossil fuels, while releasing significant amounts of CO 2 into the atmosphere.

Energy is wasted across all layers of large-scale systems. One significant source of inefficiency is the overprovisioning of the hardware. Internet-services have typically very strict performance guarantees, defined in Service-Level Agreements, which result in overprovisioning of both the computing and the supporting equipment. The manufacturing of the extra equipment is a burden to the environment. The operation of the computing equipment at periods of low-utilization, which is the majority of the time, wastes a lot of energy because server HW tends to be energy-inefficient at low utilizations. Another significant source of inefficiency is the software, which is not designed both to maximize energy-savings and optimize performance.

In datacenter planning, industry experts utilize methodologies to calculate the total cost/profit. We propose to include the environmental cost in current methodologies. With this, we demonstrate that large installations can be more energy-efficient compared to smaller ones. To tackle the cluster overprovisioning problem, researchers have proposed both HW and SW solutions. In HW, researchers have proposed the usage of low-power servers which transition all the server components into a low-power state fast. In SW, algorithms which consolidate the load into a subset of servers at periods of low utilization have been proposed. Unfortunately these solutions do not allow access to the memory during load consolidation, even though memory is a critical resource for the performance of internet-services. We propose both a memory-active, low-power family of server states and cluster middleware software which enables the management of the memory during consolidation. Our energy-efficient SW and HW achieve up to 38% energy savings in a datacenter cluster running an internet-service, over an energy-oblivious cluster. At the server level, dynamic power management solutions for the CPU and the DRAM may degrade performance with application complexities, such as highly dynamic loads or strided memory access patterns. We propose an SLA-aware power governor for the CPU and a locality-aware power management technique for the DRAM. Our CPU technique saves up to 8% system energy over the default power governor of the Linux kernel for a highly dynamic enterprise workload, while our DRAM technique improves the ED2 by 21% on average for an enterprise and scientific suite of benchmarks over a static-only power-optimized memory system.

Overall, this thesis demonstrates that large-scale systems have the capacibility to be more energy-efficient, so long that energy-efficiency is integrated as a design goal throughout the large-scale system stack.