S.S. Manvi, G.

Krishna Shyam / Journal of Network and Computer Applications ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 9

et al. (2011) propose an efficient dynamic task scheduling scheme EC2 given in Bhowmik et al. (2010) is a public cloud which
for virtualized data centers. Considering the availability and provides computing resources to general public on pay-per-use
responsiveness performance, the general model of the task sche- model. Zhang et al. (2011) showed that by strategically co-locating
duling for virtual data centers is built and formulated as a two- network I/O applications together, considerable performance gain
objective optimization. A graceful fuzzy prediction method is could be obtained. However, they did not show how to utilize this
given to model the uncertain workload and the vague availability strategy to help decision making in the cloud.
of virtualized server nodes, by using fuzzy logic systems. Kim et al. (2011) have presented a vision for the creation of
There are a few published papers on cloud computing perfor- global cloud exchange for trading services. Chabarek et al. (2010)
mance prediction model. For instance, Vianna (2012) proposed an describe mechanisms that automatically allocates service resources
analytical model to predict performance for a Hadoop online suitable for mobile devices in cloud computing environment
prototype using intra-job pipeline parallelism with no reference supporting social media services. The model is able to recommend
to power consumption. efficient virtualization by analyzing user context and the state of
Xie (2010) focuses on the optimization of the MapReduce system. In addition, this model analyzes social media service
performance in heterogeneous Hadoop clusters. The work shows resource in real time, learning user context for virtualization.
performance improvements for placing data across multiple nodes In virtualized data centers, VMs often communicate with each
so that each node has a balanced data processing performance. other by establishing virtual network topologies. However, due to
But, it does not provide a prediction model to verify and estimate VM migrations or a non-optimized allocation, the communicating
performance variations for different disks and processor architec- VMs may end up hosted on logically distant physical nodes
tures. The work does not analyze disk, I/O latency variation for providing costly data transfers between each other. If the com-
different patterns, nor does it show any improvement in the power municating VMs are allocated to the hosts in different racks or
consumption associated with the proposed optimized data placing enclosures, the network communication may involve additional
method. network switches and links, which consume significant amount of
A summary of some of the resource provisioning schemes is energy as per Chiaraviglio and Matta (2011).
given in Table 4. Table 5 lists out the performance metrics of the There have been recent research efforts on the optimization in
resource provisioning schemes. allocation of communicating applications to minimize the network
data transfer overhead as in Chiaraviglio and Matta (2010b).
However these works have not directly addressed the problem
5.1.1. Open challenges in resource provisioning
of energy consumption by the network infrastructure. Moreover,
The challenges in resource provisioning are as follows.
the proposed approaches do not optimize the placement of VMs at
 How to make the applications hosted on the cloud to be elastic run-time depending on the current network load, which is
effective for variable communication patterns and should be
so that we can achieve economy of scale while preserving the
applied to virtualized data centers.
application-specific Service Level Agreements (SLAs) such as,
In tenant-based resource allocation model, Batini et al. (2011)
response time, throughput?
 recommend some work to be done to improve and continue
How do we develop resource prediction models for facilitating
validating the infrastructure. It is recommendable to deploy a
proactive scaling in the cloud so that hosted applications are
different platform over the cloud infrastructure, such as High-
able to withstand the variation in workload with least drop in
Performance Computing (HPC) or scenarios such as online trans-
performance and availability?
 actional applications. In Upton (2010), the resource allocation is
How resources may be provisioned to an application mix such
entirely done by an online algorithm that is based on profiling
that the SLAs of all applications are met?
 active and idle time periods of desktop activity. On the other hand,
How to design resource provisioning algorithm that correctly
in Mei et al. (2010), the resource allocation on a virtual desktop is
converges to the optimal CPU allocation based on the data
entirely done by an offline algorithm that is based on resource
arrival rate and computational needs ?
 predictions from profiling user workloads in traditional desktops.
How to design a system to support n-tier clustered applications
In Morikawa and Ikebe (2011), authors propose a dynamic
hosted on a cloud ?
 resource allocation method based on the load of VMs on IaaS,
How to extend the prediction model, which is currently only
abbreviated as DAIaS. This method enables users to dynamically
used to retract over-provisioned resources, to also perform
add and/or delete one or more instances on the basis of the load
bottleneck prediction in advance, in order to overcome the
and the conditions specified by the user. They have implemented a
virtual machine boot-up latency problem ?
prototype to evaluate the effectiveness and efficiency of DAIaS.
Furthermore, they have performed an experiment to extract the
prototype on a real cloud service, namely, Amazon EC2.
5.2. Resource allocation In He et al. (2011) , authors propose an efficient resource
management solution specially designed for helping small and
Resource allocation has a significant impact in cloud comput- medium sized IaaS cloud providers to better utilise their hard-
ing, especially in pay-per-use deployments where the number of ware resources with minimum operational cost. Such an opti-
resources are charged to application providers. The issue here is to mised resource utilization is achieved by a well-designed
allocate proper resources to perform the computation with mini- underlying hardware infrastructure, an efficient resource sche-
mal time and infrastructure cost. Proper resources are to be duling algorithm and a set of migrating operations of VMs.
selected for specific applications in IaaS. Once the required types Ishakian and Sweha (2010) consider the case of a single cloud
of resources are determined, instances of these resources are provider and address the question of how to best match customer
allocated to execute the task. Resource determination and alloca- demand in terms of both supply and price in order to maximize
tion for each atomic task is managed by task modules. the providers revenue and customer satisfactions while mini-
IaaS cloud allocates resources to competing requests based on mizing energy cost.
pre-defined resource allocation policies. Presently, most of the In conventional congestion control, even when only a specific
cloud providers rely on simple resource allocation policies like resource type is congested, use of all resource types is restricted.
immediate and best effort as in http://aws.amazon.com. Amazon This brings down the efficiency in the use of other resource types,

Please cite this article as: Manvi SS, Krishna Shyam G. Resource management for Infrastructure as a Service (IaaS) in cloud computing:
A survey. Journal of Network and Computer Applications (2013), http://dx.doi.org/10.1016/j.jnca.2013.10.004i