Computational Intelligence - May 2015 - 55

Another approach to scheduling consists in first clustering the tasks, and
then assigning the clusters to the processors [11]. Such simplification leads
to the mapping problem, which can be
reduced to the bin-packing problem,
which is known to be NP-hard [12].
Bin-packing problems have online and
offline variations. The latter solves the
problem task by task, while the former
solves the problem in batches. VM colocation can be used to reduce total
resource consumption [13].
This survey section underlines the
most important characteristics of CC
workloads and systems. A set of relevant
CC resource management problems will
then be proposed and classified using
these characteristics.
B. Workload Models

CC workloads are submitted by cloud
users, i.e. individuals or organizations
that use cloud services. The tasks may
require multiple resource types and
additional objects, such as input or output data. The granularity of a task may
vary from a simple HTTP request, to
the execution of a larger process up to
the whole life cycle of a VM.
In the most simple case, the tasks are
independent, i.e. the success and finish
time of each task are accounted separately. More complex models involve related tasks. This typically implies that the
system response time is defined as the
finish time of the last task in a group.
Two such models are: the Bag of Tasks
(BoT) model, in which a set of tasks has
no additional constraints; and the Directed Acyclic Graph (DAG) model featuring precedence constrained tasks [14],
[15]. A special subset of workflow applications are multi-tier applications, which
involve simple and repetitive patterns of
dependencies between tasks, representing different layers of Web services [16].
The behavior of users, or groups of
users, can vary over time. This aspect is
noticeable at the higher CC layers as
changes in users' requests [17] and at the
lower CC layers as variable resource utilization [18], [19]. To summarize, cloud
usage patterns are heterogeneous due to
the varying user needs.

C. System Models

Table 2 Classification of cloud resource management objectives.

In the context of
CC, the notion of a ObjecTive grOup ObjecTives
resPonse Time, UPTime, ThroUghPUT,
machine (or a pro- Performance
makesPan
cessor) covers a
financial
Price, income, cosT
wide range of conenvironmenT
energy, Peak Power, Thermal, CO2 emissions
cepts and differs
oThers
reliabiliTy, secUriTy, DaTa cenTer locaTion
between studies. It
may stand not only
for a physical server,
which increases problem complexity.
but also for a VM or even a whole data
Geographic location is linked to the
center.VMs are treated as processors or as
availability of auxiliary resources (elecpieces of a workload depending on the
trical energy and its source, cool air and
layers of the cloud resource management
water) and the existence of laws or regproblem involved. Some models and
ulations that can result in restrictions on
studies propose a multi-layer approach,
the ways data are processed. If multiple
where tasks are allocated to VMs, which
data centers are run by different providare further allocated to servers [20], [21].
ers, the system can be called a multiProcessors in a system may be homogecloud. If the providers cooperate to
nous, i.e. have identical characteristics, or
provide a common platform, the system
heterogeneous. In the latter case, the exeis called a cloud federation.
cution time of a task depends of the allocated machine, as machines vary in
capabilities and speed.
D. Objectives
New or standardized data centers
The standard objectives of resource
may be composed of homogeneous
management in IT systems (summarized
machines. As time passes, and new hardin Table 2) are the response time for a
ware is shipped to data centers, they are
user, and the throughput for a service
either totally refurbished or, more often,
provider. These objectives can be further
become heterogeneous. Machines can
defined using the notions of makespan,
be described by the resource types they
which denotes the maximum compleprovision and their location in the data
tion time for a batch of tasks, and the
center network.
mean flowtime, which is defined as the
Data center networks are crucial for
mean time spent by tasks in the system.
ensuring CC performance. Important
SLAs often promote service uptime as a
considerations include network topology,
critical performance metric, i.e. the perlink latency and bandwidth. Because of
centage of time that the service is availtheir properties, links can be modeled as a
able to its users. Uptime can be related
set of machines [22]. Communication
to the success rate, which is the ratio of
activities may then be directly modeled,
the number of successfully executed
e.g. as a separate class of tasks in DAGs
tasks to the total number of tasks.
[23]. Some studies neglect the impact of
The service-related aspects of CC
networking, ignoring transmission time,
offer a second group of objectives,
and the delays and congestion that may
which are cost-related. As in other ecooccur. While peer-to-peer concepts are
nomic settings, cloud customers are
seldom explicitly mentioned in the coninterested in obtaining the best perfortext of CC, a peer-to-peer network
mance for the lowest price. On the
topology is often used to express the conother hand, the providers' goal is to
nectivity between data centers. Howmaximize their income, by maximizing
ever, some more realistic models involve
revenue and minimizing costs.
consideration of network topology,
One of the emerging objectives is the
including the modeling of switches and
minimization of the environmental
routers as networking nodes.
impact, driven by the policies of governSpatially-distributed cloud systems
ments and companies, the technical chalare composed of multiple data centers,
lenges in reliable provisioning of high

may 2015 | IEEE ComputatIonal IntEllIgEnCE magazInE

Table of Contents for the Digital Edition of Computational Intelligence - May 2015