blue / papers.iiswc2015-frontiers · Commits

Commit 0a83da7a5ece7d8c06575837c662be48f8d605ad

Authored by U-COE\mhempstead
1 parent 8327ccb6

slight changes to algorithms

Inline Side-by-side
Showing 1 changed file with 8 additions and 8 deletions
performance_clusters.tex
  View file @0a83da7
... ... @@ -46,10 +46,10 @@ sensitivity of performance clusters to number of frequency settings available in
46 46 the system.
47 47  
48 48 \subsection{Performance Clusters}
49   -We search for the performance clusters using an algorithm that is similar to the algorithm we used to find the optimal settings. We
50   -first filter the settings that fall within given inefficiency budget, and
51   -then search the optimal settings in the first pass. In the second pass, we find all of the
52   -settings that have speedup within the \textit{cluster threshold} of the optimal performance settings.
  49 +We search for the performance clusters using an algorithm that is similar to the approach we used to find the optimal settings. We
  50 +first filter the settings that fall within a given inefficiency budget, and
  51 +then search for the optimal settings in the first pass. In the second pass, we find all of the
  52 +settings that have a speedup within the specified \textit{cluster threshold} of the optimal performance.
53 53  
54 54 \begin{figure*}[t]
55 55 \begin{subfigure}[t]{\textwidth}
... ... @@ -87,7 +87,7 @@ our benchmarks, we observed that the maximum achievable inefficiency is anywhere
87 87 chose inefficiency budgets of 1 and 1.3 to cover low and mid inefficiency
88 88 budgets. %, as energy distribution among components becomes critical to extract best performance.
89 89 Cluster thresholds of 1\% and
90   -5\% allow us to model the two extremes of performance degradation bounds.
  90 +5\% allow us to model the two extremes of tolerable performance degradation bounds.
91 91 A cluster threshold of less than 1\% may limit the ability to tune less often.
92 92 While cluster thresholds greater than 5\% are probably not realistic as user is already
93 93 compromising performance by setting low inefficiency budgets to save energy.
... ... @@ -121,7 +121,7 @@ Not all of the stable regions increase in length with increasing inefficiency bu
121 121 %Increase in the length of stable regions with increase in
122 122 %inefficiency is a
123 123 %function of workload characteristics.
124   -If the consecutive
  124 +If consecutive
125 125 samples of a workload have a small difference in performance but differ significantly in energy
126 126 consumption then only at
127 127 higher inefficiency budgets will the system find common settings for these
... ... @@ -129,8 +129,8 @@ consecutive samples. % because all settings under an inefficiency budget are con
129 129 %Note that we find the performance clusters by considering
130 130 %the settings that fall under a given inefficiency budget.
131 131 This is because,
132   -performance clusters of higher inefficiencies can include settings operating at
133   -lower inefficiencies as long as their performance degradation is within set
  132 +the performance clusters of higher inefficiencies can include settings operating at
  133 +lower inefficiencies as long as their performance degradation is within the
134 134 cluster threshold. For example, the memory frequency oscillates for samples
135 135 32-39 for \textit{gobmk}
136 136 at inefficiency budget of 1.0, while the system could stay fixed at 800MHz memory at inefficiency
... ...