Reducing Operational Costs through Consolidation
with Resource Prediction in the Cloud
Jian Li
*
, Kai Shuang
*
, Sen Su
*
, Qingjia Huang
*
, Peng Xu
*
, Xiang Cheng
*
, and Jie Wang
+
*
State Key Lab of Networking and Switching Technology,
Beijing University of Posts and Telecommunications
{upton, shuangk, susen, QJHuang, xupeng, chengxiang}@bupt.edu.cn
+
University of Massachusetts Lowell
wang@cs.uml.edu
Abstract—How to achieve energy efficiency to run a cloud data
center is a major challenge in the era of rising electricity cost and
environmental protection. Various techniques have been devised
to help reduce energy consumption for cloud data centers that
consist of a large number of identical servers, including dynamic
allocation of active servers, consolidating diverse applications to
run on them, and adjusting the CPU speed of an active server.
Leveraging these techniques, we use an Online Coloring Bin
Packing problem to model the consolidation problem and devise
an effective application-aware approximation algorithm to find a
near-optimal solution. We show a 1.7 asymptotic approximation
ratio. We then apply a Predictive Bayesian Network model to
identify daily workload patterns and adjust resource provisioning
accordingly. We evaluate our approaches using traces collected
from a real data center and demonstrate that (1) our prediction
algorithm is effective in estimating future demands, (2) our coor-
dinated approaches can provide significant savings of energy and
operational costs close to the near-optimal offline solution,and (3)
our approaches incur little reliability costs in term of wear-and-
tear of server components.
Keywords: Data Center; Energy Efficiency; Consolidation;
Forecast-based Resource Provisioning
I. I NTRODUCTION
People today are increasingly relying on cloud services
to conduct business and take care of their household and
entertainment needs. Search engines, web mail, and online
social networks for examples, are common cloud services. To
guarantee performance at large scales, cloud service providers
would operate data centers to take on these IT demands. Rising
electricity cost and concerns over greenhouse gas emissions
amplify the importance of power management of these centers.
Even in 2006, data centers consumed an estimated 61 billion
kilowatt-hours (kWh) of energy, and accounted for 1.5% of
the total U.S. power consumption [1]. A recent study estimates
that operational costs for powering and cooling these centers
will soon exceed acquisition costs [2]. Thus, how to reduce
operational costs for cloud data centers has become a serious
issue.
A number of power management mechanisms have been
used in practice to reduce energy consumption. At the hard-
ware level, the Dynamic Voltage and Frequency Scaling
(DVFS) technology scales processor clock frequencies and
supplies needed voltages. At the software level, consoli-
dating applications to run on a smaller number of servers
can reduce power consumption. In this paper, we present
a new consolidation algorithm to effectively combine het-
erogeneous applications based on correlations between them,
and a forecast-based resource provisioning algorithm to better
allocate needed resources. Our approach is motivated by
observing real data collected in the Web caching project [3].
First, we observe that the increased diversity of applications
is a new trend in the cloud. Second, the total workload in
cloud data centers typically exhibits periodic patterns with
daily, weekly, and seasonal cycles [4]. The first observation
motivates consolidation of various types of workloads on a
smaller number of servers. The second observation implies
that adjusting resource provisioning during off-peak periods
can reduce power consumption.
In this context, we first formulate the consolidation problem
as an Online Coloring Bin Packing problem (OCBP). This
problem, however, is NP-hard. We then devise an Online
Coloring First-Fit (OCFF) algorithm to approximate OCBP,
which is based on First-Fit bin-packing heuristic for its
simplicity and effectiveness. OCFF automatically packs the
arrival workload into a smaller number of servers. We show
that OCFF provides a 1.7 performance guarantee. Second,
taking advantage of the periodic patterns of workload traces,
we devise a forecast-based resource provisioning algorithm
using the Bayesian Network (BN) model, which enables us
to dynamically allocate proper amount of resources (e.g., the
number of servers and their respective frequencies) and hence
reduce power consumption. Our coordinated approaches run
periodically to account for time-varying behavior. Due to
unforeseen factors such as flash crowds, we reserve sufficient
capacity readily. The main contributions of this paper are
summarized as following:
• We devise an efficient online algorithm for packing het-
erogeneous applications into a fewer number of servers.
We prove its asymptotical worst-case performance upper
bound. We also incorporate a forecast-based resource
provisioning algorithm to minimize the operational costs.
• We carry out extensive numerical experiments and
demonstrate that our approaches can significantly reduce
operational costs. We show that using online algorithm
2012 12th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing
978-0-7695-4691-9/12 $26.00 © 2012 IEEE
DOI 10.1109/CCGrid.2012.50
793