Page 86 - Proceedings of the 2018 ITU Kaleidoscope

‎ 2018 ITU Kaleidoscope Academic Conference‎

Considering the different approaches presented in Table • RQ 1: which ML techniques could be considered more
1, recent research work by the authors [14] validate that appropriate for VMPr Triggering methods?
algorithms considering prediction-based VMPr triggering
methods for solving a VMP problem in a two-phase • RQ 2: how important is to accurately predict when to
optimization scheme outperform other evaluated alternatives. trigger a VMPr phase in VMP problems?
Consequently, designing novel prediction-based VMPr • RQ 3: rather than predicting future objective function
triggering methods may improve operations of cloud values, what other parameters could be evaluated for
computing data centers, where ML techniques represent a VMPr Triggering methods?
promising approach.
Additionally to the potential of applying ML techniques for 3.2 ML for Network Management
VMPr triggering methods, several other issues should be In the context of 5G networks, mobile operations should
considered during the reconfiguration time, once a placement include critical and fault-tolerant services, where live
reconfiguration is accepted. migration of VMs between PMs in VMPr phases may require
In the next section, two main identified opportunities for short-time re-routing strategies and adaptive topologies,
ML techniques are summarized in order to improve cloud where SDN represents a valid approach in cloud computing
computing data center management. networks. Consequently, considering network routing
reconfiguration (NRR) [16] as part of VMP problems also
3. MACHINE LEARNING OPPORTUNITIES represent opportunities for ML techniques to predictively
According to Mitchell et al. [15], a broadly accepted reconfigure networking topologies and routes in cloud
definition of algorithms in machine learning fields is that: services.
A computer program is said to learn from experience E with Additionally, several formulations consider inter-VM network
respect to some class of tasks T and performance measure P traffic minimization by locating VMs with high network
if its performance at tasks in T, as measured by P, improves communication rate in the same PM, as studied in [12]. In
with experience E. this case, studying techniques for clustering these VMs may
Additionally, it is important to remember that machine result in being useful for this particular operational decision
learning may be broadly classified into [23]: making. Consequently, the following research questions may
• Supervised Learning: where inputs and desired outputs be analyzed:
(labels) are given to the learning algorithm as a training • RQ 4: which ML techniques could be considered
phase. more appropriate for predicting Network Routing
• Unsupervised Learning: where no labels are given, Reconfiguration (NRR) as part of VMP problems in SDN
leaving the learning algorithm to find structure in its implementations?
inputs. • RQ 5: which ML techniques could be considered
more appropriate for clustering VMs for supporting
In the context of the research topic presented in this work, placement decisions?
most important applications include Regression (supervised
learning) and Clustering (unsupervised learning). Section It is important to note that several other challenges and
3.1 discusses opportunities for ML techniques in Regression opportunities may be considered for applying ML techniques
and its application to prediction problems, particularly to improve cloud computing data center management in 5G
for proposing novel VMPr Triggering, while Section 3.2 service operations.
describes ML opportunities in Clustering and its application
to cloud computing network management. 4. CONCLUSIONS AND FUTURE DIRECTIONS

3.1 ML for VMPr Triggering In the context of resource management for cloud computing
In the context of the studied two-phase optimization scheme data centers, several challenges may be addressed by
for VMP problems in cloud computing environments, considering state-of-the-art machine learning techniques.
prediction-based VMPr triggering methods represent a This paper presented identified opportunities on improving
promising triggering approach, as previously studied by the critical resource management decisions, analyzing the
authors in [14]. To the best of the author’s knowledge, potential of applying machine learning to solve these relevant
existing VMPr triggering methods consider basic statistical problems, mainly in Regression and Clustering applications.
techniques (e.g. double exponential smoothing) for deciding A two-phase optimization scheme for VMP problems
when or under what circumstances to trigger a VMPr phase. was considered to present opportunities for machine
A regression analysis based on current operational data on learning techniques as a promising approach to address
cloud computing data centers may improve prediction models identified research challenges such as proposing novel
in real-world implementations, as 5G service providers. prediction-based VMPr triggering methods (see Section 3.1)
Consequently, exploring alternative techniques for predicting and applying clustering algorithms to identify VMs with high
when to trigger a VMPr phase should advance the field of the communication rates to allocate them in the same PM if
studied VMP problems, considering the following research possible in order to minimize inter-VM network traffic (see
questions: Section 3.2).

– 70 –