CRUZ, B. M. J.; CRUZ, BRUNA.; CRUZ, BRUNA M. J.; http://lattes.cnpq.br/9528471266015253; VILAR, Bruna Maria Justino Cruz.
Resumen:
The constant advances in hardware and software technologies allowed mobile computing
systems to become increasingly necessary in our everyday lives. The development of increas-
ingly complex applications for these computer systems, such as video editing and gaming,
requires more processing power. The increase in number and complexity of these appli-
cations coupled with the need for connectivity by means of some network interface means
higher power consumption. However, mobile devices have a limited power source, so efficient
use of energy is a determining factor for a good user experience.
The Dynamic Power Management (DPM) is an effective way to reduce power consump-
tion and is widely used in real systems. The most common action in power management
is the shutdown of idle devices. Although the turn off action indicates a reduction in con-
sumption, the turn on power consumption is often high enough to justify that not always
it is the best option in an idle interval to change the state of consumption of the that are
not being used effectively.
In this doctoral work, a policy for dynamic energy management, named the Optimal
Shutdown Time Policy (TOD) is developed. The workload is modeled as a two state Markov
chain, active and idle. It is considered in this work that there is an optimal time for which
the device with dynamic power management should remain off or in a low power state.
However, this time is not necessarily the entire idle interval, since the lengths of the idle
intervals vary over time. Analytical models are developed for performance penalty and for
energy consumption. An optimization problem is formulated and the policy table is defined.
So, for each known workload, the optimal shutdown time, T, is determined based on energy
saving metrics and performance penalty. At runtime, the workload model is identified
and, using a policy table, it is defined the optimal T for this situation, which will allow
predictive wake-up, minimizing energy consumption and meeting a certain performance
penalty constraint.