Page 83 - ITU Journal Future and evolving technologies – Volume 2 (2021), Issue 2

ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 2

additionally include external factors like the device’s ment of the tasks partitioning decision, i.e the decision to
residual battery [59] and so adjust to the devices’ needs process on which nodes each part of the task, can reduce

in real time. When MEC servers are UAVs, the hover time the latency. Indeed, if the decision is taken on the reques-
is to be included in the energy model under the form: ℎ ting node, it can take its much constrained resources. But

= ⋅ , where P is the power to hover and T the if the decision is taken on a remote MEC server, it may
hovering time [92]. Besides the hovering time, UAV take more time to reach other MEC servers, depending on

consumes energy for lying, depending on its velocity their placement from the device.

and weight [86]. Its accelerations have equally
signi icant impact on energy [78]. We can ignore some 4.3.3 Reliability

energy consumption points in the optimization whether
As seen in Section 3, some tasks of mission‑critical appli‑
they are idle energy and we cannot control it. This is the

cation are of vital importance. Thus, the MEC must have

case for server idle energy consumption or energy
a certain level of reliability to ensure that these tasks are
consumption of links that are traf ic independent [60,
processed. In wireless networks, the reliability is seen as
58]. Plus, some actions are negligible in comparison to

the probability to successfully transfer data within a de‑
others in the system, like downloading energy
consumption [53]. lay [93]. A irst challenge in MEC networks is node failure.
The redundancy of tasks is a relevant solution to mitigate
4.3.2 Latency this effect [94, 95]. However, it can burden the network

if the redundancy takes more than the needed compu-

Latency is crucial in mission‑critical applications where ting or communication resources. A node failure

situations may be life or death, like in search and res‑ measurement helps ensure the minimum tasks’

cue. A task’s latency comprises the processing time and reliability, avoiding the resources’ overuse [77]. Another
necessary transmission time from the device to the edge challenge is extreme events in server and UE processing

and potentially to the cloud [79, 68, 54, 65]. Work [52] queues. When queues are overloaded they may drop

add to it the compression time, present in system with some critical tasks, and assuring an average queuing

heavy tasks like video processing. We can also add the delay is not suf icient to prevent that [96]. Thus, the

local or remote computational queuing delay [55, 64] work [64] uses the statistics of the extreme queue length

because of the continue tasks generation, present even to ensure reliability.

when other tasks are processed. The processing time de‑

pends on CPU cycles required to complete the task and 4.4 Methods

the computing capacities, e.g., CPU cycles/seconds, allo‑
The chosen method for resource allocation has to propose
cated to the task [74, 69]. The latency is equally affected
a satisfactory compromise between precision, computa‑

by the data generation speed. When the generation is

tional complexity and scalability depending on the prob‑
superior to the system processing capacities, data accu‑

lem and its context. Some methods may be unable to solve
mulates in buffers and nodes don’t process tasks in real

a problem [56] or ill the system requirements. In addi‑
time. Wang et al. [61] refer to it as a blocking state and

tion, the method has to it the scale of the system, not
propose to adapt the resource allocation scheme depen-
ding on whether the system is in a blocking state or being too complex for large‑scale systems, and its needs,

in a nonblocking state. Furthermore, the data for example if suboptimal results are suf icient.
generation is usually non‑uniform across the system. It 4.4.1 Optimization methods
leads vary‑ ing workloads between servers, and some
may be over‑ loaded while others are free from tasks. It Classic mathematical optimization methods aims to solve
is so interest‑ ing to consider balancing in the resource problems optimally. Cao et al. [47] solve optimally the

allocation [69], as well as the trade‑off between resource allocation in a three‑node network to minimize
computing and transmis‑ sion time when moving a task devices’ energy consumption with the Lagrange duality

to a less loaded but further node [65]. In addition, some method. Chen et al. [68] propose a scheme for resource

devices can process critical tasks or occupy a pivotal role allocation and task placement in ultra‑dense networks

in the system, therefore they need priority in their for minimizing the task completion time. They resolve

processing. A solution proposed in [52] is minimizing a the computational resource allocation part of the prob‑

weighted‑sum delay of all devices, the weight re lecting lem with Karush–Kuhn–Tucker (KKT) conditions. Ren

devices’ importance in the system. Alternatively, [51] et al. [52] exploit the KKT conditions to allocate a MEC

proposes to measure each tasks’ pri‑ ority with delay server’s resources to users while minimizing the delay,

and reliability requirements. Standardly, the where data is compressed locally by the user before

downloading time from server to devices is ignored, sending. Even though classic mathematical optimization

since results data are smaller and downlinks have higher methods allow optimal outcomes they come with signi i‑

rates [62, 58]. What’s more, the transmission time be‑ cant complexity. Thus they are adapted to small‑scale sys‑
tween a base station and its associated MEC server is ig‑ tems with few parameters. They are unadapted to large‑

nored [59]. Finally, as seen in Section 4.1, the partition scale systems where the complexity is too high to handle

of tasks can greatly reduce the processing time by paral‑ and they will either not be able to solve the problem or
leling the processing. [45] shows that the dynamic place‑ demand an unfeasible amount of time.
© International Telecommunication Union, 2021 69