conclusions.tex
1.32 KB
\section{Conclusion}
\label{sec-conclusions}
In this work, we introduced the \textit{inefficiency} metric that can
be used to express amount of battery life that the user is willing to
sacrifice to improve performance. We used DVFS for the CPU and DFS for
the memory as a means to trade-off performance and save energy
consumption. We demonstrated that, while individual performance-energy
trade-offs of single components are intuitive, the interplay of just
these two components on the energy and performance of applications is
complex.
%We made an observation that higher energy doesn't
%always mean higher performance and vice-versa.
Consequently, we characterized the
optimal CPU and memory frequency settings across applications for multiple inefficiency
budgets. We demonstrated that if the user is willing to sacrifice minimal
performance under a given inefficiency budget, frequent tuning of the system can
be avoided and the overhead of energy management algorithms can be mitigated.
As future work, we are working towards developing predictive models for
performance and energy that consider cross-component interactions. We are
designing algorithms that use these models for tuning systems at runtime.
Eventually, we plan on designing a full-system that is capable of tuning
multiple components simultaneously while executing applications.