System-Level Power Management And An Overview:Conclusions

Conclusions

This chapter describes various DPM approaches for performing energy-efficient computation: predictive shutdown, Markovian decision process-based, and generalized stochastic Petri net-based approaches. A significant reduction in power consumption can be obtained by employing these DPM techniques. For example, for applications where continuous computation is not being performed, an aggressive shutdown strategy based on an online predictive technique can reduce the power consumption by a large factor compared to the straightforward conventional schemes where the power-down decision in based solely on a predetermined idle time threshold. Moreover, predictive shutdown heuristic may be applied to manage the shutdown of peripherals such as disks. An online algorithm that makes the shutdown decision using a prediction of the time to next disk access can result in higher power reduction compared to more conventional threshold-based policies for disk shutdown.

In contrast, construction of optimal power management policies for low-power system is a critical issue that cannot be addressed by using common sense and heuristic solutions such as those used in predictive shutdown schemes. Stochastic models provide a mathematical framework for the formulation of power-managed devices and workloads. The constrained policy optimization problem can be solved exactly in this modeling framework. Policy optimization can be cast into a linear programming problem and solved in polynomial time by efficient interior point algorithms. Moreover, trade-off curves of power versus performance can be computed. Furthermore, adaptive algorithms can compute optimal policies in systems where workloads are highly nonstationary and the service provider model changes over time.

CTMDP-based techniques introduce a new and more complete model of the system components as well as the model of the whole system. This mathematical framework captures the characteristics of the real applications more accurately which is mainly because the problem is solved in continuous-time domain while previous approaches solve the problem in discrete-time domain.

A shortcoming of DTMDP or CTMDP-based techniques is that it is very difficult to use these modeling frameworks when attempting to represent complex systems, which in turn consist of multiple closely inter- acting SPs and must cope with complicated synchronization schemes. In this case, GSPN and the correspond- ing modeling techniques based on the theory of GSPN have proven to be quite effective. The constructed GSPN model can be automatically converted to an isomorphic continuous-time Markov decision process. From the corresponding Markov decision process, one can calculate the optimal DPM policy, which achieves minimum power consumption for given delay constraints. In real applications, the interarrival time of service requests may not follow an exponential distribution, for example, they could have heavy-tail distributions such as Pareto distribution. This problem can be solved by using the “stage method” (i.e., approximating the given source of requests by a series–parallel connection of exponentially distributed sources).