**4. Results**

12 Will-be-set-by-IN-TECH

The goal of data analysis is to apply appropriate statistical tests to reject the null hypothesis. The analysis will be conducted separately for each scenario in order to evaluate which one has

In order to extract a Power Consumption profile for each Usage Scenario, a set of descriptive statistics was derived from the experimental data. For a single scenario, a total of 30 runs were executed, each composed of 300 observations (one per second) of the power consumption value. Thus, the calculations for the descriptive statistics were made using two approaches: firstly, the average of each run is extracted, obtaining a short vector of 30 elements, which was used as the subject of our analysis. This method allowed to speed up the calculations, and because of the decreased sampling rate, the data was less variant and showed an almost

Afterwards, the same analysis on the full datasets was applied, which means a total of 9000 observations. Comparing the results from these two approaches, focusing on the Index of Dispersion and the variance, the variability of a single scenario can be appreciated, which

First of all, the null hypothesis *H*10 will be tested for each scenario. Then the scenarios will be

First of all, data distribution must be analysed, in order to determine the appropriate testing method for each hypothesis. The data distribution analysis was conducted using the Shapiro-Wilk normality test. Since its results pointed out that the data was not normally distributed, non parametric tests were adopted, in particular the Mann-Whitney test [7] for testing *H*20, and the Spearman's rank correlation coefficient (also known as Spearman's *ρ*) for

The first hypothesis *H*10 is clearly directional, thus the one-tailed variant of the test will be applied. The second and third hypotheses *H*20, *H*30 are not directional, therefore the

We will draw conclusions from our tests based on a significance level *α* = 0.05, that is we accept a 5% risk of type I error – i.e. rejecting the null hypothesis when it is actually true. Moreover, since we perform multiple tests on the same data – precisely twice: first overall and then by category – we apply the Bonferroni correction to the significance level and we actually compare the test results versus a *α<sup>B</sup>* = 0.05/2 = 0.025. As regards Spearman's *ρ* significance, using 298 degrees of freedom (since 300 observations per scenario are available) the significance level of the *ρ* coefficient is 0.113. Thus, correlations coefficients resulting higher than this value can be considered as significant positive or negative correlations.

The threats of experiment validity can be classified in two categories: **internal** threats, derived from treatments and instrumentation, and **external** threats, that regard the generalization of

There are three main internal threats that can affect this analysis. The first concerns the *measurement sampling*: measurements were taken with a sampling rate of 1 second. This

**3.6. Analysis methodology**

regular distribution.

testing *H*30.

**3.7. Validity evaluation**

the work.

an actual impact on power consumption.

was also a useful tool for validating the experiment.

two-sided variant of the tests will be applied.

grouped into categories and *H*20 will be tested for each category.

#### **4.1. Preliminary data analysis**

We present in Table 4 and Table 5 the following descriptive statistics about measurements for each scenario. Tables reports in this order mean (Watts), median (Watts), standard error on the mean, 95% confidence interval of the mean, variance, standard deviation (*σ*), variation coefficient (the standard deviation divided by the mean), index of dispersion (variance-to-mean ratio, VMR).

Power consumptions show an excursion of about 11 W for both PCs, even if the baseline is quite different (an average of 87 W in Idle scenario for the Old PC, 51 W for the New PC). Moreover, the very low variability indexes ensure that the different samples for each scenario are homogeneous.

#### **4.2. Hypothesis testing**

The results of hypotheses testing of the research questions are exposed in this section.

The testing of hypothesis *H*<sup>1</sup> and *H*<sup>2</sup> are exposed in Table 6 and 7. These table report the scenarios tested, the p-value of Mann-Whitney test and the estimated difference of the medians between Idle scenario and the other ones.

Figure 5 represents the bar plot of the power consumption increase (in watts), with respect to idle, of each scenario. Figure 6 shows the box plot of scenario categories for each PC. As regards hypothesis *H*3, which evaluates correlations between resource usage and power consumption, more steps are needed. First of all, Table 8 reports the results of the Data Distribution Analysis. Then, in Table 9 and Table 10, are presented the results of the correlation


*Old-Gen PC New-Gen PC*

Energy Effi ciency in the ICT - Pro ling Power Consumption in Desktop Computer Systems 367

*Old-Gen PC New-Gen PC*

**Scenario Comparison p-value Est. Diff p-value Est. Diff** 0 - Idle vs. 1 - Web Navigation < 0.0001 -1.87 < 0.0001 -2.60 0 - Idle vs. 2 - E-Mail < 0.0001 -0.52 < 0.0001 -2.10 0 - Idle vs. 3 - Productivity Suite < 0.0001 -2.71 < 0.0001 -1.50 0 - Idle vs. 4 - IO Operation (Disk) < 0.0001 -10.41 < 0.0001 -10.80 0 - Idle vs. 5 - IO Operation (USB) < 0.0001 -10.41 < 0.0001 -10.60 0 - Idle vs. 6 - Image Browsing < 0.0001 -4.69 < 0.0001 -1.20 0 - Idle vs. 7 - Skype Call (No Video) < 0.0001 -5.10 < 0.0001 -5.00 0 - Idle vs. 8 - Skype Call (Video) < 0.0001 -9.05 < 0.0001 -11.50 0 - Idle vs. 9 - Audio Playback < 0.0001 -1.25 < 0.0001 -1.50 0 - Idle vs. 10 - Video Playback < 0.0001 -1.87 < 0.0001 -2.80 0 - Idle vs. 11 - Peer-to-Peer Data Exchange < 0.0001 -1.66 < 0.0001 -3.30

**Figure 5.** Bar Plot of per-scenario Power Consumption increase with respect to Idle

*4.2.3. Question 3: What is the relationship between usage and power consumption?*

*H*3*<sup>a</sup>* : *max*[*ρ*(*ICPU*, *P*), *ρ*(*IMemory*, *P*), *ρ*(*IDisk*, *P*), *ρ*(*INetwork*, *P*)] > *β*

**Scenario Comparison p-value Est. Diff p-value Est. Diff** Idle vs. Network < 0.0001 -2.08 < 0.0001 -3.20 Idle vs. Productivity < 0.0001 -2.71 < 0.0001 -1.50 Idle vs. File System < 0.0001 -10.41 < 0.0001 -10.60 Idle vs. Multimedia < 0.0001 -1.67 < 0.0001 -1.60 Network vs. Productivity < 0.0001 -0.31 < 0.0001 1.70 Network vs. File System < 0.0001 -6.97 < 0.0001 -6.80 Network vs. Multimedia < 0.0001 0.31 < 0.0001 1.60 Productivity vs. File System < 0.0001 -6.87 < 0.0001 -9.10 Productivity vs. Multimedia < 0.0001 0.73 < 0.0001 -0.20 File System vs. Multimedia < 0.0001 8.53 < 0.0001 8.60

**Table 6.** Hypotheses *H*1 Test Results

**Table 7.** Hypothesis *H*2 Test Results

14 Will-be-set-by-IN-TECH 366 Energy Effi ciency – The Innovative Ways for Smart Energy, the Future Towards Modern Utilities Energy Efficiency in the ICT - Profiling Power Consumption in Desktop Computer Systems <sup>15</sup>

**Table 4.** Scenarios Statistics Overview: Old-Generation PC


**Table 5.** Scenarios Statistics Overview: New-Generation PC

test using Spearman's method, with a 95% confidence interval, applied to every couple (*watt*, *variable*) for each scenario. Only the significant coefficients are listed.

*4.2.1. Question 1: Does software impact power consumption?*

*H*1 : *Pidle* �= *Pn*∀*n* ∈ [1, 11].

*4.2.2. Question 2: Is it possible to classify software usage scenarios basing upon power consumption? H*2 : *Pidle* �= *Pnet* �= *Pprod* �= *Pfile* �= *PMM*


366 Energy Effi ciency – The Innovative Ways for Smart Energy, the Future Towards Modern Utilities Energy Efficiency in the ICT - Profiling Power Consumption in Desktop Computer Systems <sup>15</sup> Energy Effi ciency in the ICT - Pro ling Power Consumption in Desktop Computer Systems 367

**Table 6.** Hypotheses *H*1 Test Results

14 Will-be-set-by-IN-TECH

**0 - Idle** 86.81 86.69 0.007 0.013 0.424 0.650 0.007 0.005 **1 - Web** 89.09 88.57 0.011 0.022 3.372 1.836 0.021 0.038 **2 - E-Mail** 88.03 87.11 0.024 0.047 5.195 2.279 0.026 0.059 **3 - Prod** 90.12 89.40 0.025 0.500 5.862 2.421 0.027 0.065 **4 - Disk** 94.12 97.21 0.048 0.095 21.12 4.595 0.049 0.224 **5 - USB** 96.41 97.10 0.024 0.046 5.047 2.246 0.023 0.052 **6 - Image** 91.97 91.48 0.041 0.081 15.474 3.934 0.043 0.168 **7 - Skype** 91.87 91.69 0.015 0.029 1.981 1.407 0.015 0.022 **8 - SkypeV** 95.40 95.75 0.020 0.040 3.844 1.960 0.020 0.040 **9 - Audio** 88.14 87.94 0.013 0.025 1.429 1.195 0.013 0.016 **10 - Video** 88.61 88.57 0.009 0.017 0.677 0.823 0.009 0.008 **11 - P2P** 88.46 88.25 0.010 0.019 0.842 0.917 0.010 0.009

**Table 4.** Scenarios Statistics Overview: Old-Generation PC

**Table 5.** Scenarios Statistics Overview: New-Generation PC

*4.2.1. Question 1: Does software impact power consumption?*

*H*1 : *Pidle* �= *Pn*∀*n* ∈ [1, 11].

*H*2 : *Pidle* �= *Pnet* �= *Pprod* �= *Pfile* �= *PMM*

**Old-Generation PC Mean Median S.E. C.I. Variance** *σ* **Var.Co. VMR**

**New-Generation PC Mean Median S.E. C.I. Variance** *σ* **Var.Co. VMR**

**0 - Idle** 51.39 51.20 0.007 0.015 0.507 0.712 0.013 0.009 **1 - Web** 54.05 53.90 0.014 0.028 1.883 1.372 0.025 0.035 **2 - E-Mail** 53.40 53.40 0.011 0.021 1.123 1.059 0.019 0.021 **3 - Prod** 53.09 52.70 0.016 0.032 2.369 1.539 0.029 0.044 **4 - Disk** 60.24 62.10 0.037 0.072 12.38 3.518 0.058 0.205 **5 - USB** 61.29 61.90 0.023 0.046 4.901 2.214 0.036 0.080 **6 - Image** 52.75 52.50 0.011 0.023 1.214 1.102 0.021 0.023 **7 - Skype** 56.23 56.30 0.016 0.032 2.420 1.555 0.027 0.043 **8 - SkypeV** 62.13 62.90 0.036 0.070 11.428 3.380 0.054 0.184 **9 - Audio** 52.87 52.70 0.006 0.012 0.315 0.561 0.010 0.006 **10 - Video** 54.14 54.00 0.007 0.013 0.420 0.648 0.012 0.008 **11 - P2P** 54.32 54.50 0.008 0.016 0.609 0.780 0.014 0.011

test using Spearman's method, with a 95% confidence interval, applied to every couple

*4.2.2. Question 2: Is it possible to classify software usage scenarios basing upon power consumption?*

(*watt*, *variable*) for each scenario. Only the significant coefficients are listed.

**Figure 5.** Bar Plot of per-scenario Power Consumption increase with respect to Idle


#### **Table 7.** Hypothesis *H*2 Test Results

*4.2.3. Question 3: What is the relationship between usage and power consumption? H*3*<sup>a</sup>* : *max*[*ρ*(*ICPU*, *P*), *ρ*(*IMemory*, *P*), *ρ*(*IDisk*, *P*), *ρ*(*INetwork*, *P*)] > *β*

**New-Generation PC Scenario Title Variable p-value** *ρ* **R2** 2 - E-Mail CPUUserTime. < 0.0001 0.42 17 % 2 - E-Mail CPUPrivTime. < 0.0001 0.43 18 % 3 - Productivity Suite CPUUserTime. < 0.0001 0.33 11 % 4 - IO Operation (Disk) PhysicalDiskTransfers < 0.0001 0.45 20 % 4 - IO Operation (Disk) LogicalDiskTransfers < 0.0001 0.45 20 % 4 - IO Operation (Disk) MemoryPages < 0.0001 0.44 19 % 4 - IO Operation (Disk) MemoryKByteAvailable < 0.0001 -0.54 29 % 4 - IO Operation (Disk) CPUC3Time. < 0.0001 -0.59 35 % 4 - IO Operation (Disk) CPUTime. < 0.0001 0.55 31 % 4 - IO Operation (Disk) CPUUserTime. < 0.0001 0.58 34 % 4 - IO Operation (Disk) CPUPrivTime. < 0.0001 0.39 15 % 6 - Image Browsing CPUUserTime. < 0.0001 0.34 12 % 7 - Skype Call (no video) NetworkPkts < 0.0001 0.62 39 % 7 - Skype Call (no video) MemoryKByteAvailable < 0.0001 -0.45 20 % 7 - Skype Call (no video) CPUC3Time. < 0.0001 -0.66 43 % 7 - Skype Call (no video) CPUTime. < 0.0001 0.52 27 % 7 - Skype Call (no video) CPUUserTime. < 0.0001 0.63 39 % 8 - Skype Call (Video) NetworkPkts < 0.0001 0.67 46 % 8 - Skype Call (Video) MemoryKByteAvailable < 0.0001 -0.62 39 % 8 - Skype Call (Video) CPUC3Time. < 0.0001 -0.88 77 % 8 - Skype Call (Video) CPUTime. < 0.0001 0.87 76 % 8 - Skype Call (Video) CPUUserTime. < 0.0001 0.9 81 % 9 - Audio Playback MemoryKByteAvailable < 0.0001 -0.34 12 % 11 - Peer-to-peer NetworkPkts < 0.0001 0.45 20 % 11 - Peer-to-peer MemoryKByteAvailable < 0.0001 -0.42 18 % 11 - Peer-to-peer CPUPrivTime. < 0.0001 0.35 12 %

Energy Effi ciency in the ICT - Pro ling Power Consumption in Desktop Computer Systems 369

**Table 10.** Spearman's *ρ* Coefficient between Power and Resource variables

**5.1. Question 1: Does software impact power consumption?**

**power consumption?**

As observed in Table 6, in both our test machines, every usage scenario consumes more power than the Idle scenario. This difference is even more evident in the New-Generation PC, where we obtain our highest increase percentage (up to 20%), as can be observed in Figure 7.

**5.2. Question 2: Is it possible to classify software usage scenarios basing upon**

As regards our second RQ, scenarios classification, results are not homogeneous: for instance, in Figure 6 it can be observed that Network category has a very wide range if compared to the others. Moreover, the comparison not always gives a clear distinction between the profiles. This suggests that a classification based on functionality can be inadequate for power consumption. Another classification may arise from the analysis of every single scenario. As can be seen from Tables 4, 5 and 6, the most power-consuming scenarios are those that involve File System, followed by Skype (both with and without Video Enabled) and Image Browsing. From the hardware point of view, these scenarios are also the most expensive in terms of system resources. Thus, classifying our scenarios basing upon resource utilization can be a more accurate way to estimate their power consumption. For instance, the power

**Figure 6.** Box Plots of Scenario Categories


**Table 8.** Data Distribution Analysis


**Table 9.** Spearman's *ρ* Coefficient between Power and Resource variables

#### **5. Discussion**

The collected data shows several facts, and gives the answers for the Research Questions.


**Table 10.** Spearman's *ρ* Coefficient between Power and Resource variables

16 Will-be-set-by-IN-TECH

**Scenario Data Distr. Max p-val. Data Distr. Max p-val.** 0 - Idle Not Normal 1.5e-63 Not Normal 2.2e-39 1 - Web Navigation Not Normal 4.4e-36 Not Normal 1.1e-20 2 - E-Mail Not Normal 9e-73 Not Normal 1.2e-19 3 - Productivity Suite Not Normal 1e-45 Not Normal 9.4e-29 4 - IO Operation (Disk) Not Normal 1.2e-46 Not Normal 8.7e-51 5 - IO Operation (USB) Not Normal 6.4e-52 Not Normal 2.5e-29 6 - Image Browsing Not Normal 1.1e-35 Not Normal 6.7e-22 7 - Skype Call (No Video) Not Normal 8.2e-30 Not Normal 3e-67 8 - Skype Call (Video) Not Normal 1.3e-35 Not Normal 5.2e-36 9 - Audio Playback Not Normal 7.9e-54 Not Normal 5.2e-44 10 - Video Playback Not Normal 1.6e-44 Not Normal 6.6e-81 11 - Peer-to-Peer Data Exchange Not Normal 8.9e-36 Not Normal 2.2e-35

> **Old-Generation PC Scenario Title Variable p-value** *ρ* **R2** 2 - E-Mail CPUC1Time. < 0.0001 -0.36 13 % 4 - IO Operation (Disk) CPUTime. < 0.0001 0.35 12 % 4 - IO Operation (Disk) CPUC1Time. < 0.0001 -0.35 12 % 5 - IO Operation (USB) CPUTime. < 0.0001 0.47 22 % 5 - IO Operation (USB) CPUC1Time. < 0.0001 -0.47 22 % 7 - Skype Call (No Video) CPUC1Time. < 0.0001 -0.39 15 % 8 - Skype Call (Video) CPUTime. < 0.0001 0.63 40 % 8 - Skype Call (Video) CPUUserTime. < 0.0001 0.53 28 % 8 - Skype Call (Video) CPUC1Time. < 0.0001 -0.7 49 % 11 - Peer-to-Peer MemoryKByteAvailable < 0.0001 -0.34 12 %

The collected data shows several facts, and gives the answers for the Research Questions.

**Table 9.** Spearman's *ρ* Coefficient between Power and Resource variables

*Old-Gen PC New-Gen PC*

**Figure 6.** Box Plots of Scenario Categories

**Table 8.** Data Distribution Analysis

**5. Discussion**

#### **5.1. Question 1: Does software impact power consumption?**

As observed in Table 6, in both our test machines, every usage scenario consumes more power than the Idle scenario. This difference is even more evident in the New-Generation PC, where we obtain our highest increase percentage (up to 20%), as can be observed in Figure 7.

#### **5.2. Question 2: Is it possible to classify software usage scenarios basing upon power consumption?**

As regards our second RQ, scenarios classification, results are not homogeneous: for instance, in Figure 6 it can be observed that Network category has a very wide range if compared to the others. Moreover, the comparison not always gives a clear distinction between the profiles. This suggests that a classification based on functionality can be inadequate for power consumption. Another classification may arise from the analysis of every single scenario. As can be seen from Tables 4, 5 and 6, the most power-consuming scenarios are those that involve File System, followed by Skype (both with and without Video Enabled) and Image Browsing. From the hardware point of view, these scenarios are also the most expensive in terms of system resources. Thus, classifying our scenarios basing upon resource utilization can be a more accurate way to estimate their power consumption. For instance, the power

18 Will-be-set-by-IN-TECH 370 Energy Effi ciency – The Innovative Ways for Smart Energy, the Future Towards Modern Utilities Energy Efficiency in the ICT - Profiling Power Consumption in Desktop Computer Systems <sup>19</sup>

coefficients are also present in the Hard Disk Index, but only in those scenarios that, clearly, involve File System operations. This means that those resources have a greater influence upon power consumption related to the others selected for the analysis. Further researches should

Energy Effi ciency in the ICT - Pro ling Power Consumption in Desktop Computer Systems 371

As expected, power consumption has always a *negative* correlation with the time spent by the CPU in the C1 and C3 states, which are power saving, low-activity states, and with the available memory, which means that using more memory has a positive correlation with power, which is a reasonable and correct behaviour. This is also a confirmation that the

Moreover, as expected, the scenarios who exhibit higher correlations are those who use more resources, such as Skype and IO scenarios. In particular, the Skype scenario with video enabled has a strong correlation with the CPU usage, probably because the real-time video

This experiment assessed quantitatively the energetic impact of software usage. It consisted in building up common application usage scenarios (e.g.: Skype call, Web Navigation, Word writing) and executing them independently to collect power consumption data. Each single scenario introduced an overhead on power consumption, which may raise up to 20% for recent systems: if their power consumption would follow a linear composition rule, the impact

The relationship between usage and power consumption was also analysed in terms of correlation between resource usage. Although a clear linear relationship did not arise, the analysis showed that some resources drive power consumption more than others, such as memory and CPU usage. Using a precise control over how an application consumes these resources, it can be possible to predict its power consumption, thus including dedicated countermeasures in the Software Design Phase – which, by itself, is the essence

Our experiment also gives us the indication that modern desktop systems, although being more energy efficient in standby and idle states, due to their higher scalability, are even more sensible to the impact of software usage on power consumption. This indicates that research should focus on reducing this impact, as it will be always more significant as time goes by. Moreover, results set the basis for future works and research projects. A more accurate correlation analysis will be conducted, focusing on the more relevant resources and taking into account also different kinds of relationships (not just linear). Moreover, we will focus our attention on battery-powered mobile devices, where software power consumption is a key issue. Our idea is that re-factoring applications by considering a more efficient resource

utilization, the impact of software on power consumption could be easily reduced.

Giuseppe Procaccianti, Luca Ardito, Antonio Vetro' and Maurizio Morisio

elaboration makes the CPU the dominant resource for power consumption.

probably focus upon these two variables.

**6. Conclusions and future works**

could be even higher.

**Author details**

*Politecnico di Torino, Italy*

of Energy-Aware Programming.

analysis was conducted with the right premises.

**Figure 7.** Per-scenario Power Consumption increase with respect to Idle (in percentage)

consumption profile of Skype is very different (about 4-5 Watts in average) with and without enabling the Video Camera.

Another interesting question that arises from the analysis is, in case of applying these Scenarios in groups, if their power consumption would follow a linear composition rule (thus summing up the values). That is, for example, supposing a composed Usage Scenario *S* that involves a Skype Call, a Web Navigation and a Disk Operation performed simultaneously, their linear composition would give, on our Old-Gen PC, an estimated Power Consumption per second of

$$P\_{idle} + \Delta P\_S = 86.81W + 21.33W = 108.14W$$

introducing a 25% overhead on power consumption. On the New-Gen PC, the estimated Power Consumption would be

$$P\_{\text{idle}} + \Delta P\_S = 51.39W + 24.90W = 76.29W$$

which gives a 48% overhead on power consumption.

#### **5.3. Question 3: What is the relationship between usage and power consumption?**

Taking a look at the results of the correlation analysis, represented by the third Research Questions, more conclusions can be made. First of all, we can observe that the coefficients related to the New-Gen PC are higher with respect to the Old-Gen PC. This may suggest that as hardware evolves, the software usage is even more significant for determining the power consumption of the system. This assumption is confirmed by Figure 7, where we can observe that the percentage increase of the New-Gen PC is higher, in most cases, with respect to the Old-Gen.

However, it is remarkable that, for both machines, the variables that show higher correlation coefficients are undoubtedly those related to CPU Usage and Memory Usage. High coefficients are also present in the Hard Disk Index, but only in those scenarios that, clearly, involve File System operations. This means that those resources have a greater influence upon power consumption related to the others selected for the analysis. Further researches should probably focus upon these two variables.

As expected, power consumption has always a *negative* correlation with the time spent by the CPU in the C1 and C3 states, which are power saving, low-activity states, and with the available memory, which means that using more memory has a positive correlation with power, which is a reasonable and correct behaviour. This is also a confirmation that the analysis was conducted with the right premises.

Moreover, as expected, the scenarios who exhibit higher correlations are those who use more resources, such as Skype and IO scenarios. In particular, the Skype scenario with video enabled has a strong correlation with the CPU usage, probably because the real-time video elaboration makes the CPU the dominant resource for power consumption.
