moving figures around

Rizwana Begum
1 parent bf41694a
Showing 1 changed file with 83 additions and 85 deletions
performance_clusters.tex
@@ -27,16 +27,6 @@ We search for the performance clusters using an algorithm that is similar to the
 first filter the settings that fall within a given inefficiency budget and
 then search for the optimal settings in the first pass. In the second pass, we find all of the
 settings that have a speedup within the specified cluster threshold of the optimal performance.
-\begin{figure}[t]
- \centering
- \includegraphics[width=\columnwidth]{./figures/plots/496/stable_line_plots/lbm_stable_lineplot_annotated_5.pdf}
-\vspace{-0.5em}
-\caption{\textbf{Stable Regions and Transitions for \textit{lbm} with
-Threshold of 5\% and Inefficiency Budget of 1.3:} Solid lines represent the
-stable regions and vertical dashed lines mark the transitions made by
-\textit{lbm}.}
-\label{lbm-stable-line-5-annotated}
-\end{figure}
 Figures~\ref{clusters-gobmk},~\ref{clusters-milc} plot the performance
 clusters during the execution of the benchmarks \textit{gobmk} and \textit{milc}. We
@@ -49,48 +39,6 @@ Cluster thresholds of 1\% and
 A cluster threshold of less than 1\% may limit the ability to tune less often.
 While cluster thresholds greater than 5\% are probably not realistic as user is already
 compromising performance by setting low inefficiency budgets to save energy.
-\begin{figure*}[t]
- \begin{subfigure}[t]{\textwidth}
-	\centering
- \vspace{-1em}
- \includegraphics[width=\columnwidth]{{./figures/plots/496/stable_line_plots/stable_lineplot}.pdf}
- \end{subfigure}%
-\vspace{0.5em}
-\caption{\textbf{Stable Regions of \textit{gcc} and \textit{lbm} for
-Inefficiency Budget of 1.3:} Increase in cluster threshold increases the length
-of the stable regions, which eventually leads to less transitions. Higher
-inefficiency budgets allow system to run unconstrained throughout.}
-\label{stable-regions} 
-\end{figure*}
-\begin{figure*}[t]
- \begin{subfigure}[t]{\textwidth}
-	\centering
- \includegraphics[width=\columnwidth]{{./figures/plots/496/stability_metrics/transition_percent_across_thresholds}.pdf}
- \end{subfigure}
-\vspace{-0.5em}
-\caption{\textbf{Number of Transitions with Varying Inefficiency Budgets and
-Cluster Thresholds:} The number of frequency transitions decrease with increase in cluster
-threshold. The amount of change varies with benchmark and inefficiency budget.} 
-%Decrease in number of transitions is dependent on benchmark and
-%inefficiency budget.}
-\label{transitions-bar}
-\end{figure*}
-
-\begin{figure*}[t]
- \begin{subfigure}[t]{\textwidth}
-	\centering
- \includegraphics[width=\columnwidth]{{./figures/plots/496/stable_length_box/stable_length_box}.pdf}
- \end{subfigure}%
-\vspace{0.5em}
-\caption{\textbf{Distribution of Length of Stable Regions:} The average length of
-stable regions increases with cluster threshold.
-%While inefficiency budget has significant impact on length of stable regions
-%for \textit{lbm}, it doesn't change the length of stable regions of
-%\textit{gobmk} much.
-}
-\label{box-lengths}
-\end{figure*}
-
 Figures~\ref{clusters-gobmk}(c),~\ref{clusters-gobmk}(d) plot the
 performance clusters for \textit{gobmk} for inefficiency budget of 1.3 and
@@ -142,6 +90,29 @@ performance within set cluster threshold (for example samples 3-5 in
 Figures~\ref{clusters-gobmk}(a),~\ref{clusters-gobmk}(c)).
 Figure~\ref{clusters-milc} shows that \textit{milc} has similar trends as
 \textit{gobmk}.
+\begin{figure}[t]
+ \centering
+ \includegraphics[width=\columnwidth]{./figures/plots/496/stable_line_plots/lbm_stable_lineplot_annotated_5.pdf}
+\vspace{-0.5em}
+\caption{\textbf{Stable Regions and Transitions for \textit{lbm} with
+Threshold of 5\% and Inefficiency Budget of 1.3:} Solid lines represent the
+stable regions and vertical dashed lines mark the transitions made by
+\textit{lbm}.}
+\label{lbm-stable-line-5-annotated}
+\end{figure}
+\begin{figure*}[t]
+ \begin{subfigure}[t]{\textwidth}
+	\centering
+ \vspace{-1em}
+ \includegraphics[width=\columnwidth]{{./figures/plots/496/stable_line_plots/stable_lineplot}.pdf}
+ \end{subfigure}%
+\vspace{0.5em}
+\caption{\textbf{Stable Regions of \textit{gcc} and \textit{lbm} for
+Inefficiency Budget of 1.3:} Increase in cluster threshold increases the length
+of the stable regions, which eventually leads to less transitions. Higher
+inefficiency budgets allow system to run unconstrained throughout.}
+\label{stable-regions} 
+\end{figure*}
 An interesting observation from the performance clusters is that algorithms
 like CoScale~\cite{deng2012coscale} that search for the best performing settings every interval starting
@@ -150,6 +121,33 @@ overhead of optimal settings search by starting search from the settings selecte
 for the previous interval as application phases are often stable for multiple sample intervals.
 %as the application phases don't change drastically in
 %short intervals.
+\begin{figure*}[t]
+ \begin{subfigure}[t]{\textwidth}
+	\centering
+ \includegraphics[width=\columnwidth]{{./figures/plots/496/stability_metrics/transition_percent_across_thresholds}.pdf}
+ \end{subfigure}
+\vspace{-0.5em}
+\caption{\textbf{Number of Transitions with Varying Inefficiency Budgets and
+Cluster Thresholds:} The number of frequency transitions decrease with increase in cluster
+threshold. The amount of change varies with benchmark and inefficiency budget.} 
+%Decrease in number of transitions is dependent on benchmark and
+%inefficiency budget.}
+\label{transitions-bar}
+\end{figure*}
+\begin{figure*}[t]
+ \begin{subfigure}[t]{\textwidth}
+	\centering
+ \includegraphics[width=\columnwidth]{{./figures/plots/496/stable_length_box/stable_length_box}.pdf}
+ \end{subfigure}%
+\vspace{0.5em}
+\caption{\textbf{Distribution of Length of Stable Regions:} The average length of
+stable regions increases with cluster threshold.
+%While inefficiency budget has significant impact on length of stable regions
+%for \textit{lbm}, it doesn't change the length of stable regions of
+%\textit{gobmk} much.
+}
+\label{box-lengths}
+\end{figure*}
 \subsection{Stable Regions}
 So far, we have made observations by
@@ -185,39 +183,6 @@ currently designing algorithms in hardware and software that are capable of tuni
 running the application as future work. In Section~\ref{sec-algo-implications}, we
 propose ways in which length of stable regions and the available settings for a
 given region can be predicted for energy management algorithms in real systems. 
-\begin{figure}[t]
-% \begin{subfigure}[t]{0.45\textwidth}
-%	\centering
-% \includegraphics[width=\columnwidth]{{./figures/plots/496/energy_perf_bar/energy_bar_normalized_0.0_0_0}.pdf}
-% \vspace{-2.5em}
-% \end{subfigure}%
-% \begin{subfigure}[t]{0.45\textwidth}
-	\centering
- \includegraphics[width=\columnwidth]{{./figures/plots/496/energy_perf_bar/performance_bar_normalized_0.0_0_0}.pdf}
-% \vspace{-2.5em}
-% \end{subfigure}%
-\caption{\textbf{Variation of Performance with Inefficiency:} Performance
-improves with increase in inefficiency budget, but the amount of improvement varies across benchmarks.
-%\XXXnote{Is this description sufficient?}
-%System always stays under the given inefficiency budget. 
-%Maximum inefficiency consumed by the system is bounded by the application and
-%devices.
-}
-\label{ineff-perf}
-\end{figure}
-
-\begin{figure*}[t]
- \begin{subfigure}[t]{\textwidth}
-	\centering
- \includegraphics[width=\columnwidth]{{./figures/plots/496/energy_perf_bar/energy_perf_bar_1.3}.pdf}
- \end{subfigure}
-\vspace{-0.5em}
-\caption{\textbf{Energy-performance Trade-offs for Inefficiency Budget of 1.3,
-Multiple Cluster Thresholds:} Performance degradation is always within the
-cluster threshold. Allowing small degradation in performance reduces energy
-consumption, which decreases further when tuning overhead is included.}
-\label{energy-perf-trade-off}
-\end{figure*}
 Figure~\ref{stable-regions} plots stable regions for benchmarks \textit{gcc} and
 \textit{lbm} for multiple inefficiency budgets and cluster thresholds. With
@@ -258,6 +223,39 @@ Therefore the number of transition per billion instructions decrease only slight
 with increase in cluster threshold and inefficiency budget for \textit{gobmk}.
 Figure~\ref{box-lengths}(c) summarizes the distribution of stable region lengths observed
 across benchmarks for multiple cluster thresholds at inefficiency budget of 1.3.
+\begin{figure}[t]
+% \begin{subfigure}[t]{0.45\textwidth}
+%	\centering
+% \includegraphics[width=\columnwidth]{{./figures/plots/496/energy_perf_bar/energy_bar_normalized_0.0_0_0}.pdf}
+% \vspace{-2.5em}
+% \end{subfigure}%
+% \begin{subfigure}[t]{0.45\textwidth}
+	\centering
+ \includegraphics[width=\columnwidth]{{./figures/plots/496/energy_perf_bar/performance_bar_normalized_0.0_0_0}.pdf}
+% \vspace{-2.5em}
+% \end{subfigure}%
+\caption{\textbf{Variation of Performance with Inefficiency:} Performance
+improves with increase in inefficiency budget, but the amount of improvement varies across benchmarks.
+%\XXXnote{Is this description sufficient?}
+%System always stays under the given inefficiency budget. 
+%Maximum inefficiency consumed by the system is bounded by the application and
+%devices.
+}
+\label{ineff-perf}
+\end{figure}
+
+\begin{figure*}[t]
+ \begin{subfigure}[t]{\textwidth}
+	\centering
+ \includegraphics[width=\columnwidth]{{./figures/plots/496/energy_perf_bar/energy_perf_bar_1.3}.pdf}
+ \end{subfigure}
+\vspace{-0.5em}
+\caption{\textbf{Energy-performance Trade-offs for Inefficiency Budget of 1.3,
+Multiple Cluster Thresholds:} Performance degradation is always within the
+cluster threshold. Allowing small degradation in performance reduces energy
+consumption, which decreases further when tuning overhead is included.}
+\label{energy-perf-trade-off}
+\end{figure*}
 \subsection{Energy-Performance Trade-offs}
 In this subsection we analyze the energy-performance trade-offs made by our