Performance Analysis for Large IaaS Clouds - Large Scale and Big Data: Processing and Management - page 575

18.5 DISCUSSIONS AND RELATED RESEARCH
In this section, we discuss relevance of the models for Big Data environments, pos-
sible extensions of the developed models and outline related research.
18.5.1 i interaCting s ubmoDels For b ig D ata a nalytiCs
Processing large volumes of data sets typically requires IaaS type of cloud services.
In most cases, overall execution time of a Big Data processing job reduces with
larger the number of PMs and VMs. However, provisioning such VMs quickly as
the service demand increases is an equally important but less investigated issue.
This chapter shows how model-driven analysis can be performed to improve over-
all provisioning delay. Specifically, scalability of the developed approach facilitates
modeling of large IaaS pools that can host such Big Data processing environments.
18.5.2 F uture e Xtensions
Although we presented in this chapter the analysis for three pools, the proposed
approach can be extended n (>3) pools in general. For each pool, one CTMC is
needed for a VM provisioning submodel. Also, the RPDE submodel will have n
rows to capture the decision process across n pools. Thus, the developed modeling
approach scales linearly with the number of pools in the cloud.
VM provisioning submodels presented in this chapter model homogeneous PMs.
They can be extended to model heterogeneous PMs as well. Heterogeneous PMs can
be divided into multiple classes with varying characteristics and each class can be
represented by a pool. Then, PMs will be homogeneous within a pool and heteroge-
neous across the pools.
Different provisioning strategies can also be modeled using VM provisioning
submodels. In this chapter, we assumed that only one job is being provisioned at a
time, while other jobs wait in the queue. This assumption can be easily relaxed to
model parallel provisioning of the jobs. We can use a state dependent multiplier to
the provisioning rates (β h , β w , and β c ) for modeling such cases.
18.5.3 r elateD r esearCh
Xiong et al. [23] used response time distribution as a QoS metric for cloud performance
analysis. However, their stochastic models do not capture details of cloud service pro-
visioning decision, VM provisioning, and release. In [23], Varalakshmi et al. addressed
workflow scheduling in cloud to meet user-requested QoS parameters. The authors
analyzed the PM performance behaviors using Generalized Processor Sharing queues.
Mills et al. [16] used sensitivity analysis methods to identify important input
parameters that can affect cloud placement algorithms. Mi et al. [17] developed an
approach to detect performance bottlenecks in large cloud services. Using hierarchi-
cal structure, the authors constructed an execution path graph of user request. Such
an approach can be combined with our performance model to optimize the request
placement decisions.