|Thesis abstract: |
Power consumption has become a major concern for almost every digital system: from the smallest embedded circuits to the biggest computer clusters, an energy and power budget is always constraining the performance of the system. Moreover, the actual power consumption of these systems is strongly affected by their current "working state" (e.g., from idle to heavy-workload conditions, with all the shades in between), which is often a consequence of the external interactions they are subject to. Given these assumptions, it is difficult to make accurate predictions on the power consumed by the whole system over time, when it is subject to constantly changing operating conditions: this makes the definition of energy saving policies not trivial in most of real world scenarios.
Given the amount of research opportunities in the field and the limitations of the works in literature, I am currently focusing my Ph.D. on the development of an holistic power modeling tool that can be used to profile the energy consumption of a wide range of energy-constrained systems: the same high-level workflow is tailored on the actual system¿s features, extracting a specific power model able to describe and predict the future energy behavior of the observed entity. This methodology will be provided in an ¿as-a-service¿ fashion: at first, the target system is instrumented to collect power metrics and workload statistics in its real usage context; then, the collected measurements are sent to a remote server, where data is processed using well known techniques (e.g., Principal Components Analysis, Markov Decision Chains, AutoRegressive statistical models, etc.); finally, an accurate power model is built as a function of the metrics monitored on the instrumented system. The profiled system, being it an embedded device as well as a node of a multi-tenant server infrastructure, will then be able to take advantage of energy and power predictions to optimize its behavior with respect to its own performance-per-watt goals.