Page 85 - Proceedings of the 2018 ITU Kaleidoscope

Machine learning for a 5G future

The proposed formulation of the VMPr problem receives the 2.2.1 Periodical Triggering
following information as input data:
As presented in Table 1, several studied works have looked
at periodically triggering the VMPr phase. Periodically
• a set of n available PMs and their specifications; triggering the VMPr could present disadvantages when
defining a fixed reconfiguration period (e.g. every 10 time
• information about the utilization of resources of each instants). For example, a reconfiguration could be required
active VM at discrete time t; before the established time, where optimization opportunities

• the current placement at discrete time t (i.e. x(t)). could be wasted or even economical penalties could impact
cloud data center operation. On the other hand, in certain
cases the reconfiguration may not be necessary and triggering
The considered optimization scheme for the VMP problem the VMPr could represent profitless reconfigurations.
is based on methods to decide when or under what
circumstances to trigger placement reconfigurations with 2.2.2 Threshold-based Triggering
the migration of VMs between PMs (VMPr triggering) and Another regularly studied VMPr triggering method considers
what to do with cloud services requested during placement a threshold-based approach (see Table 1), where thresholds
recalculation time (VMPr recovering). The VMPr phase is are defined in terms of utilization of PM resources (e.g.
triggered according to a given VMPr triggering method (see CPU). Thresholds indicate when a PM H i is considered to be
Section 2.2). underloaded or overloaded, and consequently, a VMPr should
Once the VMPr is triggered, the placement of VMs at be triggered. For example, fixing utilization thresholds for
discrete time t is recalculated during β discrete time slots overloaded and underloaded PM detection, for all considered
(i.e. recalculation time). The result of the VMPr problem resources, to 10% and 90% respectively. The described
threshold-based VMPr triggering method makes isolated
is a placement reconfiguration for the discrete time t − β
(i.e. x (t − β)). It is important to note that the recalculated reconfiguration decisions at each PM without the complete
0
placement is potentially obsolete, considering the offline knowledge of global optimization objectives, giving place to
nature of the VMPr phase. In fact, while the VMPr a distributed decision approach.
is making its calculation, the iVMP still may receive
and serve arriving requests, making obsolete the VMPr 2.2.3 Prediction-based Triggering
calculated solution; therefore, the recalculated placement Considering the main identified issues related to the studied
must be recovered accordingly using a VMPr recovering VMPr triggering methods, prediction-based VMPr triggering
method, before complete reconfiguration is performed. The methods were recently proposed in the VMP specialized
recovering process as well as the migration of VMs are
performed in γ discrete time slots (i.e. reconfiguration time), literature, statistically analyzing an objective function F(x, t)
that is optimized and proactively detecting situations where
where γ may vary according to the maximum amount of RAM a VMPr triggering is potentially required for a placement
to be migrated. Figure 1 presents the described two-phase reconfiguration. The considered prediction-based VMPr
optimization scheme, considering β = 2, from t = 2 to t = 4 triggering method uses a double exponential smoothing
and γ = 1, from t = 4 to t = 5. (DES) [7] as a statistical technique for predicting values of the
It is important to note that a large number of possible objective function F(x, t), mathematically formulated next:
objective functions F(x, t) and constraints e(x, t) could be
considered for a VMP problem formulation, according to S t = α × Z t + (1 − τ)(S t−1 + b t−1 ) (3)
provider preferences [11, 10].
2.2 Considered VMPr Triggering Methods
b t = τ(S t − S t−1 ) + (1 − τ)(b t−1 ) (4)
A VMPr triggering method defines when or under what
circumstances a VMPr phase should be triggered in a Z t+1 = S t + b t (5)
two-phase optimization scheme for VMP problems. By where:
considering studied VMPr triggering methods (see Table α: Smoothing factor, where 0 ≤ α ≤ 1;
1), three main approaches may be identified: (1) periodical, τ: Trend factor, where 0 ≤ τ ≤ 1;
(2) threshold-based and (3) prediction-based. The following Z t : Known value of F(x, t) at discrete time t;
sub-sections describe the VMPr triggering methods evaluated S t : Expected value of F(x, t) at discrete time t;
in this work as part of a two-phase optimization scheme for b t : Trend of F(x, t) at discrete time t;
VMP problems. Z t+1 : Value of F(x, t + 1) predicted at discrete time t.
Table 1 – Summary of studied triggering methods.
At each discrete time t, the prediction-based VMPr triggering
References VMPr triggering method predicts the next M values of F(x, t) and effectively
[4, 21, 6, 9, 5, 22, 19] Periodically triggers the VMPr phase in case F(x, t) is predicted to
[2, 18, 20] Threshold-based consistently increase, considering that F(x, t) is being
[14] Prediction-based minimized.

– 69 –