Page 90 - Kaleidoscope Academic Conference Proceedings 2020

2020 ITU Kaleidoscope Academic Conference

used four hyper-parameters (criterion for split = MAE, in the Docker container virtualization platform demonstrated
learning rate = 0.3, number of estimators = 50, and the efficacy of the proposed resource control scheme. We
subsample size = 0.8) and for ETR two hyper-parameters observed that the ETR model can predict the resource
(criterion for split = MAE and number of estimators = 28). utilization of sine and Poisson workload patterns with good
accuracies, namely, mean absolute errors of 0.89% and
4.2 Resource utilization prediction with online 2.66%, respectively.
retraining
In future work, we will extend this work to applying the
The comparisons of actual and predicted resource utilization prediction of resource utilization to perform the actual
values with Poisson and sine input workload patterns (i.e., resource adjustment with the online retraining models. We
QPS) are illustrated in figures 6 and 7, respectively. The will study more about the optimal size of a training data set
findings from these figures are summarized in Table 1. and hyper-parameters tuning of ETR and GBR. The
proposed scheme is related to ML-based network control and
Figure 6 shows that the prediction performances of both management methods that are currently being standardized
GBR and ETR improved with retraining (between 100-200 in the ITU-T Study Group 13.We will also bring this research
seconds). However, applying more training data might outcome gradually to ITU for standardization. The IoT-
degrade the prediction performance by overfitting the directory service, which has been used in the performance
predictions (between 200-300 seconds). Initially, the evaluation of the proposed resource adjustment scheme, is
predicted utilization for the sine wave workload was far based on Recommendation ITU-T Y.3074. It is a scalable
behind the actual utilization. However, it converged quickly directory system that can store control information of a
after four retraining iterations (around 60 seconds in Figure billion IoT devices in the form of name records and provide
7). Due to this reason, is positive for both ETR and GBR a very fast lookup service with the latency of a few
in Table 1. For the Poisson workload, since the variation in milliseconds.
workload was less, the predictions were almost similar to or
marginally higher than the actual ones after the first round of REFERENCES
retraining (around 10 seconds in Figure 6), resulting in a
negative value of for both models. The ETR model scored [1] M. Fukuyama, “Society 5.0: Aiming for a New
the MAE values of 0.89% and 2.66% for sine and Poisson Human-Centered Society,” Japan SPOTLIGHT,
workload patterns, respectively, while the GBR model July/August 2018, pp. 47-50.
scored 1.02% and 2.85%, respectively.
[2] G. Aceto, V. Persico, and A. Pescapé, “A Survey on
There were sporadic actual high resource utilization values, Information and Communication Technologies for
specifically for the Poisson workload, which might be due to Industry 4.0: State-of-the-Art, Taxonomies,
the fluctuations in the state of an experimental system’s Perspectives, and Challenges,” IEEE Commun.
virtualization platform. To cope with this situation, we ran Surveys & Tutorials, Vol. 21, Issue 4, 2019, pp.
the experiment several times to keep the occurrences of 3467 - 3501.
sporadic high utilization values at a minimum number of
times and presented the best-case results for sine and Poisson [3] “The Tactile Internet,” ITU-T Technical Watch
workloads here. Report, Aug. 2014.

The capability of accurately predicting resource utilization [4] M. Chiang, S. Ha, C. I, F. Risso, and T. Zhang,
enables us to dynamically adjust the allocated resources and “Clarifying Fog Computing and Networking: 10
maintain the desired performance. Based on the accurate Questions and Answers,” IEEE Commun. Mag., vol.
estimation of resource utilization, we can calculate the 55, no. 4, pp. 18-20, Apr 2017.
required amount of computational resource to keep the
resource utilization value less than a threshold value (say [5] ITU-T Recommendation Y.3074, “Framework for
80%) and achieve the desired performance (i.e., lookup directory service for management of huge number of
latency kept lower than a maximum tolerable value). We can heterogeneously named objects in IMT-2020,” 2019.
use a Docker update command to adjust the computational
resource as described in [10]. [6] S. Shekhar, H. Abdel-Aziz, A. Bhattacharjee, A.
Gokhale, and X. Koutsoukos, “Performance
5. CONCLUSION Interference-Aware Vertical Elasticity for Cloud-
Hosted Latency-Sensitive Applications,” in Proc.
We have proposed and evaluated a computational resource 11th IEEE Cloud, San Francisco, CA, 2018, pp. 82-
control scheme employing multiple regression models as ML 89.
techniques. It can support in dynamic resource adjustment
decision making. Its accuracy is increased by online
retraining of the models training data set collected from the
running system. Our experimental results obtained from a
test bed implementation of the IoT directory service function

– 32 –