3.2 Energy Aware Frame Based Fair Scheduling (EAFBFS)
3.2.4 Experiment and Results
3.2.4.2 Performance Evaluation of EAFBFS Algorithm
We have evaluated our algorithm with different input parameters and compared the results for type 1 and type 2 systems. The results of the evaluation has been discussed below:
(a)Normal Set (b) Skewed Set
Figure 3.2: Effect of varying utilization factor on power consumption (n= 96, φ= 0.1|T Sr| and σwt= 0.3)
Experiment 1 (Effect of Utilization Factor): The utilization factor of a system represents the ratio of the summation of weights of the tasks and the number of available cores. In our experiments, we have varied the utilization factors between 40% and 100%.
The system under consideration has n = 96, σwt = 0.3 and φ = 0.1|T Sr| with 8 or 16
cores. As we can see from Figure 3.2, the power consumption is directly proportional to the utilization factor of the system. As the utilization factor rises, the idling time for the cores within a time-slice reduces, thus allowing less scope for lowering operating frequencies, ultimately resulting in higher power consumption. In a type 2 system, we can independently vary the operating frequency of individual cores. So, EAFBFS gives better results on a type 2 system than on a type 1 system. The average difference in power consumption between type 1 and type 2 systems were found to be 4.0% and 23.49% on 8 and 16 cores respectively (Figure 3.2a). To evaluate power consumption considering skewed task sets, the degree of skewness has been taken to be 1:4. In this case (Figure 3.2b), the relative difference between the weights of tasks is high; consequently, the frequency demand of the tasks also varies significantly. The average difference in power consumption were found to be 34.25% and 36.49% on 8 and 16 core systems, respectively. Hence, the difference in power consumption between type 1 and type 2 systems become more significant for skewed task sets. Type 2 systems adapt well to the skewness among task weights.
(a)Normal Set (b) Skewed Set
Figure 3.3: Effect of varying number of cores on power consumption (n= 96, φ= 0.1|T Sr| and σwt= 0.3)
Experiment 2 (Effect of the number of cores): We evaluated our algorithm on systems having 4 to 28 cores with utilization factors of 50% and 70%. The rest of the system specifications are same as in the setup used for performance evaluation against changes in utilization factor (Experiment 1). Figure 3.3 shows that power consumption
3.2 Energy Aware Frame Based Fair Scheduling (EAFBFS)
of type 1 systems grow rapidly with increase in the number of cores, while for type 2 systems, the growth is almost negligible. When the number of cores increase in a system, for a given utilization factor and number of tasks, the average individual task weights in- crease. Thus, we have more cases where the individual frequency demand (Equation 3.9) of a task is higher than the average system frequency (Equation 3.8). It may be observed that as the weight of a task increases in a system, the individual frequency demand also increases, which in turn results in increased power consumption. This leads to more wastage of power in a type 1 system, as all cores in a type 1 system are restricted to operate at the same frequency. Correspondingly, this effect is far milder in case oftype 2 systems as they are able to maintain an optimal frequency on all the cores. Thus the increase in number of cores has less impact on power consumption. In particular, when the number of cores is varied from 4 to 28 with 50% system utilization, the normalized power consumption in type 2 systems remain almost same while it increases by 68.55%
on type 1 systems.
The average difference in power consumption between type 1 and type 2 systems for normal task sets having utilization factors 50% and 70% respectively, were found to be 31.26% and 28.44% (Figure 3.3a). The differences in power consumptions between type 1 and type 2 systems become more prominent when the task weights are skewed with respect to each other. This has been revealed in Figure 3.3b, where the values of the difference in power consumption between type 1 and type 2 systems for 50% and 70% utilization factors are 50.22% and 45.72%, respectively.
Experiment 3 (Effect of Skewness): We followed two approaches to evaluate and compare the power consumption of EAFBFS against skewness of a task set. In the first approach, we considered 16 core systems having utilization factors of 50% and 70%, with n = 96 and φ = 0.1|T Sr|. We varied the standard deviation of the distribution for generating task weights from σwt = 0.1 to σwt = 0.5, keeping (µwt) = 0.5 as con- stant. From Figure 3.4a, we observe that as the standard deviation among task weights increase, type 2 systems are able to provide higher energy savings as compared totype 1 systems. This may be attributed to the fact that increase in standard deviation in turn
(a) Effect of variation in standard deviation of task weight
(b)Effect of variation in skewness within Task Set
Figure 3.4: Effect of Skewness on Normalized Power Consumption
increases the skewness among task weights, thereby also increasing the skewness among the summation of task weights within each individual core inside a time-slice. As a type 2 system allows cores to run at mutually distinct frequencies, such skewness among the summation of task weights in different cores provide type 2 systems with a better opportunity of extracting possible energy savings while meeting the fairness and dead- line constraints. We observe from Figure 3.4a that as the utilization factor of the system increases from 50% to 70%, the difference in power consumption between type 1 and type 2 systems gets higher. This is due to the fact that higher utilization factors lead to higher average task weights which in turn lead to higher variation in frequency demand by the individual tasks. The average difference in the power dissipation between type 1 and type 2 systems was observed to be 19.48% and 32.69% for systems with utilization factor of 50% and 70%, respectively.
In the second approach, we considered 16 core systems having utilization factor of 70% with 32 and 96 tasks andφ= 0.1|T Sr|. The task weights have been generated from normal distributions with standard deviation σwt = 0.1 and mean (µwt) = 0.5. The generated task sets were divided into two subsets, the first containing 80% of the total number of tasks and the second containing the remaining 20%. For the first subset, the ratio of the summation of weights of the tasks with respect to the system’s utilization
3.2 Energy Aware Frame Based Fair Scheduling (EAFBFS)
factor is varied from 1:5 to 4:5. Naturally, the ratio varies from 4:5 to 1:5 in the second subset. From Figure 3.4b, we observe that power consumption on type 2 systems only varies slightly with respect to the variation in skewness. However, the variation is significant for type 1 systems. It may be noted that for this task distribution, the task weights will be maximally uniform with least skewness when the ratio is 5:1. Hence, the power consumption of the EAFBFS algorithm in a type 1system has been observed to be lowest at point 80:20 (as shown in Figure 3.4b). As the uniformity among task weights decrease (i.e. skewness increase), the resultant power consumption rises intype 1 systems. We further observe that power consumption in type 1 systems is higher when the number of tasks is low (n= 32) than when it is high (n= 96). This is because as the number of tasks is increased while keeping the utilization factor constant, the average weight of individual tasks decrease. Thus, we have less cases where the individual frequency demand (Equation 3.9) of a task is higher than the average system frequency (Equation 3.8). In particular, when skewness is raised from 80:20 to 20:80 in type 1 system with n = 32, the power consumption increased from 81.26% to 94.49%. The average difference in the consumption of power betweentype 1 and type 2 systems with n = 32 and n= 96 was found to be 45.71% and 36.46% respectively.