UC San Diego

U.S. data center energy consumption is expected to rise past 100 billion kWh in 2013. Approximately 50% of this energy usage can be attributed to servers, networks, and storage, and the other half goes to power and cooling infrastructures in their support. Servers are so non- energy efficient that they consume 65% of a fully utilized server's power when only 30% utilized. Even when 100% utilized, the server may not be running efficiently. This dissertation improves the energy efficiency of data center systems in three ways. The first method improves server energy efficiency by turning off CPU cores during long- latency memory accesses with no performance penalty to data center applications. This technique leverages peak rush current from on-chip power distribution networks to quickly charge core capacitance, and allows the core to resume execution in as little as 8.06 ns. Core state is saved through careful use of slave latches and source biasing. A key means to increasing effective CPU utilization and energy efficiency of servers, is to leverage virtual machine and thread migration at minimal performance overhead. Our second technique speeds up software thread migration by up to 2.5× compared to Linux , with latencies as small as 933 ns. We leverage this technique to quickly migrate operating system code between asymmetric cores to reduce application energy consumption. The techniques introduced so far assume that I/O devices have sufficient bandwidth to keep the server processor busy. In contrast to this assumption, many data centers oversubscribe their networks to reduce cost and power consumption, sometimes at the expense of overall data center efficiency. Our last contribution is a software top -of-the-rack switch capable of offloading unmodified TCP/ IP traffic onto a prototype, microsecond optical circuit switch within a microsecond. We demonstrate that servers can utilize up to 95.4% of optical circuit bandwidth even when switch reconfiguration latency is reduced by three orders of magnitude to 11.5 microseconds, supporting the introduction of low-latency optics into the data center to radically reduce cost and power consumption of full bisection bandwidth networks