Performance Analysis for Large IaaS Clouds - Large Scale and Big Data: Processing and Management - page 557

user signs up for the service contract. We observe that, to date, most of the IaaS
cloud providers offer SLAs only in terms of guaranteed availability. With increas-
ing demand and popularity of cloud services, we believe that performance SLAs
will also be necessary in the near future. However, performance analysis of a cloud
infrastructure is difficult due to a variety of reasons. Hardware (e.g., CPU speed,
disk properties), software (e.g., nature of hypervisor), workload (e.g., arrival rate),
and many other management characteristics (e.g., placement policy) can impact the
overall cloud performance.
To evaluate the performance of a cloud, broadly three choices can be made. First,
one can carry out experimentation for measurement based performance quantifica-
tion. Unfortunately, scale of cloud becomes prohibitive in terms of time and cost of
such measurement-based analysis. Second, discrete event simulation can be used as
another alternative [1]. Still, such simulation can take a long time to get statistically
significant results. Third, stochastic models can be used as a low-cost option where
the model solution time is much less compared with simulation and experimenta-
tion. However, stochastic models may not scale given the size and complexity of a
cloud system. For cloud environments processing large data sets, the state space of
the developed Markov model tends to be very large as the model takes into account
many details of the system. This well-known largeness problem of stochastic model
arises because of a growing number of model states as the number of system com-
ponent increases. As the model size becomes prohibitively large, the generation and
solution of such a model become a tedious task, if not impossible. Simplifying a
model to reduce complexity may turn out to be fatal as the model might diverge
from the realistic scenario. Hence, a scalable modeling approach that can preserve
accuracy is of interest.
In this chapter, we describe the principle of scalable stochastic modeling approach
with an example of performance analysis for an IaaS cloud that are large scale and
well-suited for Big Data environments. The proposed approach is based on interac-
tions among several submodels, where the overall solution is composed by iteration
over individual submodel solutions. Scalability and tractability are two key features
of our approach when compared with a one-level monolithic modeling approach
[8]. Thus, interacting submodels can provide results for large clouds within rea-
sonable solution time. A comparison of analytic-numeric results between the two
approaches also show the accuracy of the proposed interacting submodels approach.
An additional benefit of the proposed approach is that submodels often become sim-
ple enough to obtain a closed-form solution. Stochastic modeling software packages
such as SHARPE [20] and SPNP [11] can be complemented using such closed-form
expressions when the system becomes too large.
Key contributions of this chapter are the following: (1) We demonstrate that scal-
able stochastic models can be developed and solved for large-scale performance
analysis of IaaS clouds. (2) Interested readers can extend submodels shown in the
chapter by modifying the provided SHARPE codes. (3) Such modeling approaches
can help in what-if analysis of overall cloud performance. Especially, we show how
bottlenecks in provisioning delay can shift with varying cloud capacity and job ser-
vice characteristics.