UC San Diego
Adaptive architectures for peak power management
- Author(s): Kontorinis, Vasileios
- et al.
Power budgeting - allocating power among computing components to maximize performance - is a necessity for both modern CPUs and systems in general. Our industry expects every new chip generation to double performance. However, non-ideal transistor scaling makes it hard to provide the expected performance at the same power envelope. The available power budget dictates microprocessor performance and functionality. At the system level, power provisioning affects infrastructure and energy costs. Power associated costs like wall-power cost, cooling, and power delivery have already surpassed the server costs in state-of-the-art data centers. Every component of the computing stack is associated with a power budget, and its respective marginal increase in performance must justify the respective power increase. Power budgeting is a difficult issue to address. Different workloads stress different computing resources and even a single application may stress different resources at different points of time. Thus, a single universal solution is hard to achieve. Currently, most systems fix the decision of what resources consume more power at design time. Instead, in this thesis we propose to dynamically customize resources to boost performance while respecting power budgets. This thesis presents techniques that manage power in three domains: the processor core, 3D -die stacked chips, and the data center. First, we describe a microarchitectural solution that always keeps a portion of the core powered-off, but what gets powered off can change for each application. By intelligently providing the most critical resources, we achieve a particular peak power goal while minimizing performance impact. Second, we dynamically reconfigure the resources of 3D-stacked cores. Adaptation in this context allocates unused back-end resources of one core to an adjacent core in the z-axis that needs them. Hence, we can use lower- power cores and boost their performance. Finally, we present a solution to reduce data center costs by using the energy stored in UPSs (Uninterruptible Power Supply). By discharging UPS batteries during high utilization and recharging them during low utilization, we adapt available power capacity. This approach permits more functionality under the provisioned power and can translate to cost savings of millions of dollars