Documentos de Académico
Documentos de Profesional
Documentos de Cultura
Abstract—The cost of electricity contributes significantly to the operating expense incurred in hosting cloud services. It is necessary
to consider this cost while charging the consumers for their service utilization. In this work, we arrive at a metering mechanism for
cloud services, in which the price of a cloud service tracks the variable input cost of electricity from a smart grid. The power-aware
cloud metering developed here is a dynamic pricing and billing model where tariff for a cloud service is varied in accordance with the
input electricity cost. We arrive at a model for power consumption of virtual machines hosted on the cloud infrastructure. This power
consumption model is used in calculating the cost of operation of the service. A cloud instance leased by a consumer is billed based
on the cost of operation obtained, and its resource utilization. Experimental results validate the approach presented.
Index Terms—cloud computing, smart grid, power-aware metering, billing, resource management, power consumption regulation
1 I NTRODUCTION the mapping between the input electricity cost and the
output price of the cloud service. The problem is of
the variability of electricity price across geographies to tariffs are implemented in various power grids across
reduce data center costs. Our work extends the idea to the globe [29].
vary the charge on the consumers based on the variation Electricity pricing provided by smart meters [11] in
in electricity costs. the power domain provides a starting point for us
In this paper we describe a power-aware metering to price cloud services. The power-aware metering
mechanism for cloud services. This approach caters to proposed in this paper aims at providing a dynamic
the need for a dynamic pay-per-use model in the cloud pricing mechanism for cloud service providers. The
environment based on real-time pricing of electricity. service provider specifies a price based on the current
We use the analytical power model proposed by Heath electricity cost input from the power grid and cloud
et al. [20] to arrive at the power consumption of a utilization using the power model described. It is then
cloud instance hosted on the infrastructure based on its at the consumer’s discretion to continue using the
resource utilization. The premise of this work is to use cloud instance by paying the specified amount, or to
the variability in cost of electricity to run the data center terminate the cloud instance. We propose to achieve
as a basis to determine the price of the cloud service. To a metering mechanism analogous to the advanced
this end, this model enables us to record the resource metering infrastructure in the power grids.
usage per cloud instance and to generate a bill based on The computing resource needs of a cloud consumer
two factors, viz., are time-variant, and hence the power footprint of the
consumer also varies with time. It is well known that the
(i) the marginal power consumption of the cloud
cost of power is the primary cost of maintaining a large
instance; and
data center and that this cost is passed on to consumers.
(ii) the current input cost of electricity as specified by
Given the reality of smart metering of power supplies by
the smart power grid/source.
electricity suppliers the world over, it therefore makes
Studies show that the energy consumption rises and falls sense for the vendor to use similar smart metering
with computing resource utilization [4], [5], [8], [17]. This technology to recover the cost of power from consumers.
is one of the bases for our approach, where we use the It is also in the best interest of the consumer to have a
resource utilization to determine the power consumption varying service price, as then the consumer can decide
of the cloud instance. on service utilization based on the cost. There is a need
to consider these facts while metering service utilization.
Davison et al. [21], in the course of developing a Current implementations of metering mechanisms [30],
model for electricity spot prices, indicate that the [31], [32] in cloud environments are based on service
price of electricity is seasonally variant. Power price level agreements (SLA) or bid amounts which govern
is guided by demand and supply factors. Alvarado the provisioned capacity. These are fixed-cost, post-paid
and Rajaraman [22] discuss price volatility in electricity service models. The limitation of these models is that
markets and reasons for the prevalent spot prices. consumers are billed based on a predefined tariff for
Domingues [23] sheds light on the impact of renewable the utilization. This predefined tariff does not consider
energy on the pricing of electricity in a liberalized the variation in the input cost of electricity. There is
market. It is also known that solar energy usage leads to no existing system that performs metering of a cloud
variable power price [24], because, after all, solar energy service with due consideration to the varying cost of
supply itself is inherently variable [25]. The same is also power consumption. This is our motivation to develop
true of wind energy supply [26]. Thus, as solar and power-aware metering for cloud services.
other alternative-energy sources see increasing use, we
may expect that variable power supply cost will become We propose a dynamic pricing scheme for cloud services
an inevitable reality for all large power consumers which varies according to the input cost of electricity.
including IT systems used by cloud providers. Specifically, we do the following in this paper:
(i) Classify the existing power models for servers as
Smart meters in the power grid, also called advanced
component-specific power models and system-level
metering infrastructures (AMI) [27], are devices that
power models, and compare them.
include basic hourly interval meters, meters with
(ii) Arrive at a power consumption model for cloud
one-way communication, and real-time meters with
instances based on a system-level power model.
built-in two-way communication capable of recording
(iii) Develop a cost model for cloud services which
and transmitting instantaneous data. The recorded data
is aligned with the dynamic electricity pricing
are provided to both consumers and energy companies
provided by a smart grid.
at least once daily [28]. AMI help realize time-based rates
(iv) Use the above to generate the dynamic price
and demand-response actions at the user end. These
charged for a cloud service based on the input cost
devices perform real-time registrations of electricity
of electricity.
usage, whose data are used for billing and other
purposes. They allow obtaining the usage information This work has not been done in any of the earlier
both locally and remotely. Smart metering and smart papers such as [33] which look at using electricity smart
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON SERVICES COMPUTING
3
Figure 2.
3.1 Component-Specific Power Models
3 P OWER M ODEL E VALUATION Lefurgy et al. [36] for example study power consumption
The power consumption of a subsystem of a working of memory and microprocessor subsystems of
system in production environments cannot be monitored commercial servers. Energy consumption is based
using instrumented hardware. This is because such on the structured workloads and there is a system-wide
instrumentation results in reduction of operational approach to reduce the power consumption. Bianchini
efficiency of the system, and might also interfere with and Rajamony [37] provide an exhaustive survey
other aspects like cooling and ease of maintenance. of the then state-of-the-art mechanisms for power
Therefore, in order to determine the power consumption consumption. These are some of the earlier attempts at
of the infrastructure components in real time, it is power modeling the data center hardware. Some of the
necessary to have a method to characterize the power more recent work can be characterized as following.
consumption. This is achieved by using a power model Ranganathan et al. [38] determine the power
which determines the system power consumption based consumption of the system by estimating the power
on certain parameters that define the system resource consumption of the CPU. The argument presented is
utilization. These parameters are collected at run time that the CPU is the greatest consumer of power in a
using software. blade server, and hence the system power consumption
We classify the power models into two major is determined using the processor power consumption
classes, viz., (i) component-specific power models; within an error bound of 10%. The approach followed
and (ii) system-level power models. by them is to determine the 90th percentile resource
utilization and multiply it by the rated power for the
Component/subsystem-specific power models system. The results obtained in their work involve
characterize the power consumption of such components significant levels of approximation.
or subsystems accurately. However, when a system
view of power consumption is considered, the error Deng et al. [39] derive a relation for memory power
increases when the power consumptions of other consumption in their work based on the number of
subsystems are not considered. Multiple such power LLC misses that occur per unit time. Memory power
models may therefore be necessary to obtain the entire consumption is determined as a sum of the power
system’s power consumption, which would increase the consumption of bus, bank and the memory chip.
complexity of the metering mechanism.
Though these component-specific models are accurate in
To circumvent the problems of loss of accuracy and their own accord, using any of these independently in
increase in complexity, we evaluate some power models our work results in extreme levels of approximation on
which provide a view of the system-level power one hand while, using several of them together would
consumption. Many power models that provide system result in increasing the complexity of the metering
power consumption are developed using instrumented mechanism. Hence we evaluate the possibility of using
hardware. Owing to the fact that it is impossible to a system level power model, as indicated in the next
deploy instrumented hardware in production systems subsection.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON SERVICES COMPUTING
5
Symbol Definition
Common symbols
Ptotal Total power used by the system
Pidle Power consumed when the system is idle
P Total power consumed by the cloud server farm
pj power consumed by cloud instance j
Equation (1)
Pcpu Power consumed by the CPU subsystem
Pmem Power consumed by memory subsystem
Pio Power consumed by the memory subsystem
Pnet Power consumed by the networking subsystem
Equation (3)
Pi Power consumed by the ith subsystem
fi Operating frequency of the ith subsystem
Vi Voltage used to power the ith subsystem
Equation (4)
Ppeak Power consumed when the system is at 100% utilization
Active% Activity level of the system against the full capacity
P arallelization Parallelization factor achieved in instruction execution
F etchedOps CPU performance counter which indicates the number of fetched operations per second
Cycles Number of CPU cycles per second
Equation (5)
B Power consumed when the system is idle (analogous to Pidle )
Ui Utilization of the ith subsystem
Ci Capacity of the ith subsystem
i Indicates a subsystem in the system, it is one of { cpu, memory, io, network }
TABLE 1
Symbols used in power model evaluation
TABLE 2
Comparison of power models
3.2 System-Level Power Models achieve power-aware application packing in large cluster
deployments. The goal of their work is to minimize the
There have been various approaches to power modeling power consumption in a data center. Jang et al. [41]
at a systems level. Some of these approaches use discuss energy reduction methods using memory-aware
instrumented hardware to determine the power VM scheduling. All of the above use some form of power
consumption of the component/sub-system while model to suit the need.
others determine the power consumption of the system
empirically. A few approaches under each of these In the above, the power consumed is represented as the
categories are discussed in this subsection to help sum of the subsystems’ power consumption.
appreciate the necessity of an analytical power model Ptotal = Pidle + Pcpu + Pmem + Pio + Pnet (1)
rather than using an empirically determined power
model.
A full system view of power consumption of the server
Complete system view for power consumption: is obtained by Economou et al. [42]. This model uses the
Recently, Choi et al. [40] describe a mechanism to resource utilization values of the server subsystems and
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON SERVICES COMPUTING
6
power model to characterize the power consumption of discussion and (5) that the power consumed by the cloud
the cloud instances. We derive the power consumption instance is:
of a cloud instance using an analytical model for ucpu umem
servers (5). This is a generic approach without needing pj = Mcpu × + Mmem ×
Ccpu Cmem
any instrumented hardware or calibration. uio unet
+Mio × + Mnet × (6)
Cio Cnet
where,
3.3 Cloud Instance Power Consumption
pj = power consumed by the cloud instance j in Watts,
Mi = measure of power consumption of resource i of
A cloud instance is essentially a VM hosted on one of
hosting machine full utilization (in Watts),
the bare metal servers in the data center. In order to
uj = utilization of resource i by cloud instance j, and
model the power utilization of a cloud consumer for the
Ci = capacity of resource i on the hosting machine.
service availed, it is necessary to determine the power
consumption of the particular cloud instance assigned to As the power consumption of the network and i/o
the consumer. Cloud instances are hypervisor processes subsystems are negligible when compared to the power
running on the servers in the data center. Hence to consumed by the processor and memory subsystem [43],
determine the instance power utilization, we monitor the these terms can be ignored from (6) without causing a
resource utilization of the process in which the cloud major error in the total power consumed. McCullough
instance is running. et al. determine that the power consumption of the
processor subsystem accounts for 85% of the total
There is no direct way to determine the power system power consumption and the memory subsystem
consumption of a cloud instance. The two possible ways accounts for less than 10% of the total system power
of determining the power consumption of VMs are: consumption. Based on the findings in [43], we can
(i) using instrumented hardware to run a single VM conclude that the power consumption of the memory
and determine the power consumed by the machine; subsystem is close to 11% of that of the processor
and (ii) estimating the power consumed by a VM by subsystem. Hence for all practical purposes we can
recording its resource utilization. consider the below equation to determine the power
Some earlier methods [9] use instrumented hardware consumed by the cloud instance.
to host a single VM. The power consumption of the ucpu umem
pj = Mcpu × + Mmem × (7)
server is recorded using instruments attached to the Ccpu Cmem
server. Power consumption of the server while idle is or,
also noted. The power consumed by the individual VMs ucpu
in this setup is calculated as the difference in the power pj = 1.11 Mcpu × (8)
Ccpu
consumed by the server when the VM is operational,
and the idle power consumption.
The equation (8) above represents the marginal
Chen et al. [8] provide an extensive discussion of increment in the total power consumed by running the
profiling the power consumption of VMs. The key instance j. This equation is used later to develop a cost
finding of their work is that when the vCPU is model of the data center hosting the cloud; specifically,
mapped to threads on the physical CPU, as in case to calculate the total power consumed by the hosting
of a cloud environment, the power consumption of infrastructure (9).
the VM follows the power consumption of the hosting
hardware. It is known that the kernel-based VM [47]
architecture spawns VMs as independent threads on the 4 E CONOMIC M ODEL
hosting hardware. Hence, the power consumption of the
virtualized cloud increases linearly with an increase in We establish the relation for cost of operation as a
the workload running on the cloud instances. A similar function of the price of electricity specified by the
argument is provided for the memory subsystem. Power electricity service provider, the resource utilization, and
consumption by a VM is considered to depend on the the base operational cost to keep the infrastructure
actual physical resources utilized by the VM. See [8] for running. A pricing scheme and the bill generation
details. Taking this key finding for a premise, we develop process are discussed in Section 4.2.
the following power model for the cloud instances, based
on the resource utilization of the VM.
4.1 Cost Model
Now, to determine the power model for cloud instances,
let ui be the utilization of resource i of a cloud instance j. From (8) in the previous section, we determine the
A cloud instance is essentially a process (like KVM or power consumed by a cloud instance that is deployed.
Xen) executing on a server. It follows from the above The power consumption of the cloud hosting data
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON SERVICES COMPUTING
8
Symbol Definition for pricing Internet services. However, these are not
Ct Total cost of running the cloud service at applicable in the current cloud scenario given variable
time t electricity costs.
B Base cost of running the service
Current service providers like Amazon [50] and
ξ Cost of power as indicated by the
electricity service provider
Rackspace [51] price cloud instances mostly based
Pcloud Power consumed by cloud on their configuration and duration of use. Another
p(j,t) Price of the cloud instance j at time t prevalent practice is to charge consumers a fixed price
C Total system resource capacity for a lease period like the Amazon reserved instance [50].
φ Fixed price for utilizing the cloud
Assuming that we know the capacity of the hosting
uj Resource utilization by cloud instance j
k, α Constants defined by the service
infrastructure, we divide the base cost of service among
provider all the cloud instances that are hosted. This component
forms the fixed cost of using the cloud service by
TABLE 3 consumer φ.
Symbols used in the economic model
Our pricing model incorporates the following
parameters: (i) fixed cost of utilizing the cloud
service; (ii) an elastic cost which varies according
center is the sum total of the power consumed by to the utilization; and (iii) balance factor cost which is
all cloud instances and the power needed to keep the used to reduce the gap in the available resource and
infrastructure running. Hence, the total power consumed actual resource demand.
is given by: The price per cloud instance is defined by:
Pcloud = B + pj (9)
Ct
j p(j,t) = φ + k uj × − α(C − uj ) (11)
C j
The marginal cost of operation of one cloud instance is
where,
then determined as a function of:
p(j,t) = price of the cloud instance j at time t (in $),
• the current cost of power as indicated by the C = the total system resource capacity,
electricity service provider for a defined time φ = fixed cost of utilizing the cloud
interval, ξ ($/kWh); and Ct = total cost of operation of the cloud service from (10)
• the base cost (B) incurred in running the service uj = resource utilization by cloud instance j, and
(this includes software license costs, hardware k, α are constants defined by the service provider based
maintenance cost). on the market economics. k is the scaling factor which
determines the marginal price of the cloud instance and
We define the cost of the cloud service at time t as: α is the penalty multiplier to ensure a balance between
Ct = B + ξ · Pcloud (10) the utilized and available resource. Since we do not
cover the market economics of cloud services, we do not
where, include details on how to determine these constants.
Ct = Total cost (in $) of running the cloud at time t The price per cloud instance (11) ensures the following:
B = Base cost of the service,
ξ = Cost of electricity as specified by the power service 1) The service provider is not burdened by the cost
provider in $/kWh, of running the service in terms of license fee,
Pcloud = Power consumed by cloud in Watts. maintenance etc. This cost is split equally across
the deployed cloud instances and is captured in the
We need to note that the cost of operation defined price per instance (φ).
above is variable due to amortization and increase in 2) When the resource utilization
the license refresh cost. However, in our work we do is low, the correction
factor recorded by (C − uj ) is high, hence the
not focus on these aspects and abstract these as B. The price of the instance is reduced. This encourages
cost of operation defined in (10) is used to calculate the consumers to avail the cloud service in terms of
the bill amount for the cloud instance hosted on the number of instances or increased instance capacity.
infrastructure. 3) When the resource utilization is high, the correction
factor is near zero, hence the price per instance is
4.2 Pricing Scheme the actual cost of operation. This being the case,
the consumers are discouraged to demand more
Pricing for the cloud service can be applied based on number instances and/or capacity.
multiple considerations. Buyya et al. [48] list various 4) The consumers are charged appropriately for the
pricing considerations. Historically, flat-pricing models resource utilization (via the instances owned by
and other simple proposals [49] were considered them).
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON SERVICES COMPUTING
9
The time of operation is divided into equal time intervals Algorithm 1: Resource Utilization Monitoring
called time slots for the purpose of sampling service output: vector instanceUtilizationData
utilization in the cloud infrastructure and the utilization 1 begin Instance Monitoring
by an individual consumer. The time slot granularity 2 while true do
is so defined that the price specification time of the 3 foreach Deployed Cloud Instance do
power grid is an integral multiple of the time slot 4 instanceUtilizationData ←
specified in the metering module. The other constraint RetrieveResourceUtilization(Instance)
to be considered while selecting the time slot is that 5 Store(instanceUtilizationData)
the amount of load generated on the system due to 6 end
monitoring is minimal. For instance, if T is the time 7 end
interval at which power prices are updated by the 8 end
electricity service provider, we can express the condition
above mathematically as, T = ρ · t, where ρ is some output: vector infrastructureUtilizationData
integer constant. 9 begin Infrastructure Monitoring
10 while true do
The billing amount is computed as a summation of the 11 foreach resource ∈ [cpu, memory, disk] do
instantaneous pricing obtained in (11) for each cloud 12 infrastructureUtilizationData ←
instance. The total bill amount for a cloud instance (j) RetrieveResourceUtilization(Server)
is obtained as: 13 Store(infrastructureUtilizationData)
Billj = p(j,t) (12) 14 end
t 15 end
where, p(j,t) is the operational price at time t for instance 16 end
j from (11). Data: vector
17 begin Store
18 save vector data in application database
5 I MPLEMENTATION 19 end
Resource monitoring of both the cloud instances Data: Cloud instance or Physical server
and the underlying hardware is necessary for the output: vector
metering mechanism described in the system model. 20 begin RetrieveResourceUtilization
For the purpose of monitoring the resource utilization 21 utilizationData ← get utilization data from
of the cloud instances, we can use the hypervisor performance counters or using an available tool
counters provided by the various vendors like Microsoft // this is a comment
HyperV [52]. Standalone enterprise monitoring tools return : utilizationData
22 end
like Hyperic SIGAR [53] can also be employed for
the purpose of resource usage monitoring of the cloud
instances. Any of the established enterprise data center
monitoring tools like Hyperic [54] or Nagios [55] can Algorithm 2 details the calculation of the bill for
be used to monitor the data center equipment resource every cloud instance deployed. For the calculation of
utilization. an instance level bill, the stored resource utilization
of the particular cloud instance is retrieved from
Generation of billing information follows directly from the application database (lines 10–12). The power
the cost model and the resource utilization. Once the consumption of an individual VMs is calculated using
billing interval is specified, the number of time intervals (6). The instantaneous price for the instance is calculated
of operation is determined. For each such interval, the as per (11). For the given time interval (e.g., daily,
bill amount is calculated as presented in the (11). The weekly or monthly), the total bill is calculated using
total bill is then calculated as the summation of these (12). The cost of operation of the cloud service is
values. calculated as in (10). This is internal to the calculation
of instantaneous price. The process of calculating the
Algorithm 1 depicts the resource utilization monitoring.
bill amount is repeated for all instances deployed on the
The two components of the algorithm are infrastructure
cloud infrastructure (lines 2–8).
monitoring (lines 9–16) and deployed cloud instance
monitoring (lines 1–8) respectively. Since monitoring
is a continual operation, the process is represented in 6 R ESULTS
the while block. Usage data are collected from either
the infrastructure components or the deployed cloud For the purpose of testing the model proposed, a
instances (procedure listed in lines 20–22) and saved into Eucalyptus [56] cloud was set up on x86-64 based
the application database (lines 17–19) for further use by machines running Ubuntu Linux. We used four Intel
the prediction and bill calculation modules. i7 quad core machines for node servers and one Intel
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON SERVICES COMPUTING
10
Core2Duo machine for the cloud controller and storage Fig. 4. Utilization pattern of the workloads.
controller. Another dedicated Intel Core2Duo based
desktop machine was used as a client. A setup similar
workloads on individual instances. We refer to
to the schematic representation of cloud as shown in
them as W L1, W L2, W L3 and W L4 in this work.
Figure 1 was realized.
These workloads are representative of generic cloud
Standard Ubuntu images [57] were deployed on the workload profiles: W L1 and W L3 are compute-intensive
cloud instances. The resource utilization of the cloud workloads and reach peak utilization frequently, hence
instances were recorded using off-the-shelf monitoring their power consumption is also high (results are based
tools like the Hyperic monitoring framework and on the power model described in Section 3). However
the Hyperic SIGAR framework. Cloud instances were W L2 and W L4 have low-moderate resource demand
loaded using the Apache web server [58] and MySQL mostly and have occasional spikes in the utilization.
database [59]. To induce additional load on the Their power consumption also follows a similar pattern
infrastructure, artificial load generating processes were as the resource utilization. The characteristics of the
introduced on the nodes. The configurations of the workload are depicted in Figure 4.
load-generating processes were varied to change the
A plot representing the overall system resource
infrastructure utilization as desired.
utilization is shown in Figure 5. The straight line in
As a part of the test, the input cost of electricity the graph indicates the cloud’s capacity. It was noted
was varied based on the data available from the during the evaluation that the responsiveness of the
real-time electricity supply charges provided by PJM infrastructure is within acceptable limits until the load
Interconnection LLC [60]. Variation in the electricity cost average reached a value of 3 per compute core. Hence
provided as input to the metering application is shown considering the cloud setup in our deployment had 16
in Figure 3. This graph corresponds to the electricity cores, the cloud capacity is capped at 48. When the
cost from 0800 hrs to 1900 hrs on 30 May 2012. The utilization is beyond the capacity, a penalty price is
data collection interval was 10 minutes. These data were charged to the consumer and the corresponding price
obtained by PJM [60] for the Eastern Interconnection for that peak period increases. This is captured while
Grid of the United States. calculating the instantaneous price in (11).
For purposes of evaluation, we deployed four The cost of operation represented in (10) tracks the
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON SERVICES COMPUTING
11
Fig. 5. Overall system resource utilization. Fig. 7. Variation of the service price in accordance to the
change in input cost of electricity and utilization.
input electricity cost. It also considers the total power The resource utilization of the cloud instances were
consumption of the infrastructure (Pcloud ). The variation monitored and the bill generated as per (12). Based on
in cost of operation as per input electricity cost is shown the resource utilization of the cloud instance and the total
in Figure 6. It may be noted that the two curves in utilization of infrastructure resources, a sample itemized
the graph are at different scales. The cost of the input bill per running instance in the test setup was generated.
electricity is represented in the continuous line and This bill captures the cloud instances in consideration,
the corresponding scale is the left-y-axis and the cost their power utilization and the price to be paid by the
of operation (represented by dotted lines) is according consumer who owns the cloud instances. The itemized
to the scale on the right-y-axis. One may notice that power consumption for four cloud workloads and the
there are peaks in the cost of operation between 1040hrs billed price for each of them is listed in Table 4.
to 1120hrs. This is attributable to the peaks in the
power consumption during this time period (Figure 5). Cloud Power consumed Billed amount
Similarly the valleys in the cost of operation between instance (in KWh) (in Dollars)
1500hrs to 1540hrs can be attributed to both the factors of W L1 3.21 65.76
lower electricity cost and low cloud resource utilization. W L2 0.71 15.05
It is interesting to note that between 1720hrs and 1740hrs, W L3 2.99 60.79
even though there is a valley in the electricity cost, the W L4 1.04 22.92
overall cost of operation has a peak, owing to the fact TABLE 4
that there was a peak in the cloud resource utilization Itemized listing of power consumption & billed price on
during the period. the day of observation
Pricing for the cloud service was calculated as indicated
in (11). The result of the variations in service pricing
due to varying cost of electricity were recorded. This There was a noticeable variation in the cloud instance
is represented in Figure 7. The data range represented price in accordance with the resource utilization. This is
in the graph is a snapshot of 11 hours duration from as per the expected behavior based on the cost model
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON SERVICES COMPUTING
12
7 C ONCLUSION
An effective relationship between input electricity cost,
cloud resource utilization, and individual utilization
of the cloud consumer is obtained in the proposed
model of power-aware cloud metering. Using this
relation, dynamic pricing of the cloud services based
on the utilization of the cloud infrastructure and input
electricity cost is achieved. Based on this dynamic
Fig. 9. Variation of price of the service for two workloads
pricing information, a varying tariff model for the cloud
using different power models.
services is obtained in this work.
Specifically, the following goals are met in our work.
and the pricing scheme proposed. As discussed earlier,
utilization directly affects the power consumption. • Arrive at a power consumption model for the cloud
The cost of the cloud service varies with the power instances.
consumption. The utilization data and the corresponding • Develop a cost model for cloud service which is
price of the cloud service were calculated as per the aligned with electricity pricing provided by a smart
varying electricity cost. The graph in Figure 8 represents grid.
the variation of price with utilization over a subset of • Track the price the cloud service based on the input
the entire test data. cost of electricity.
Based on the tests performed, the following can be Our metering mechanism enables cloud computing
inferred: infrastructures to directly work with smart grids by
monitoring the power usage and helping the CIO enforce
• As the input cost of electricity changes, the policies to charge consumers appropriately for the power
cost of operation of the cloud service changes consumed.
correspondingly. Hence the price charged to the
consumer varies. The impact of a high utilization on Our work thus has relevance in the context of
the infrastructure is directly observed in the output cloud services—especially infrastructure-as-a-service
price, i.e., the price levied on the customer. (IaaS)—which need to operate in the presence of
• When the utilization of the infrastructure is below a dynamic electricity prices and new and forthcoming
certain level as determined by the service provider, smart grid technologies.
there is a minimal price levied on the consumer. This An analysis of the impact of dynamic pricing can be
price accounts for the idle time of the infrastructure performed to provide a comparison between the current
as well as fixed and amortized costs. This follows service pricing mechanisms and the dynamic pricing we
directly from (10) where a base operational cost is propose.
accounted for.
• The system model described here provides for a
metering mechanism which is power-aware. R EFERENCES
In order to compare the performance of the metering [1] L. A. Barroso, “The price of performance,” ACM Queue, vol. 3,
mechanism described here, we used two different power no. 7, pp. 48–53, Sep 2005.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON SERVICES COMPUTING
13
Fig. 8. Variation of cloud service price in accordance with utilization and electricity price.
[2] P. M. Curtis, Maintaining Mission Critical Systems in a 24/7 [Online], Tech. Rep., September 2011, available at:
Environment. Wiley-IEEE Press, 2007, ISBN 978-0471683742. http://www.datacenterdynamics.com/research.
[3] K. Brill, Understanding the True Cost of [13] Y. Sverdlik, “Global data center energy use to grow by 19%
Operating a Server. facilitiesnet, Nov. 2008. [Online]. in 2012,” [Online], September 2011, available at: http://goo.gl/
Available: http://www.facilitiesnet.com/datacenters/article/ 5WnQP.
Understanding-the-True-Cost-of-Operating-a-Server--10063 [14] “Global census shows datacentre power demand grew 63% in
[4] L. Barroso and U. Holzle, “The case for energy-proportional 2012,” [Online], available at: http://goo.gl/Sszpj.
computing,” Computer, vol. 40, no. 12, pp. 33–37, Dec 2007, doi: [15] A. Verma, G. Dasgupta, T. K. Nayak, P. De, and R. Kothari, “Server
10.1109/MC.2007.443. workload analysis for power minimization using consolidation,”
[5] C. Lefurgy, X. Wang, and M. Ware, “Server-level power control,” in Proceedings of the 2009 USENIX Annual Technical Conference, ser.
in Fourth International Conference on Autonomic Computing (ICAC USENIX’09. USENIX Association, 2009, p. 28.
2007). Los Alamitos, CA, USA: IEEE Computer Society, Jun. 2007. [16] M. Uddin and A. A. Rahman, “Server consolidation: An approach
[6] N. Singh and S. Rao, “Modeling and reducing power to make data centers energy efficient and green,” CoRR, vol.
consumption in large IT systems,” in 4th Annual IEEE International abs/1010.5037, 2010.
Systems Conference (IEEE SysCon 2010), San Diego, CA, Apr. 2010, [17] D. Meisner, B. T. Gold, and T. F. Wenisch, “Powernap: Eliminating
pp. 178–183, doi:10.1109/SYSTEMS.2010.5482354. server idle power,” SIGPLAN Not., vol. 44, no. 3, pp. 205–216, Mar.
2009, doi: 10.1145/1508284.1508269.
[7] ——, “Energy optimization policies for server clusters,” in 6th
[18] A. Narayan, S. Rao, G. Ranjan, and K. Dheenadayalan, “Smart
Annual IEEE Conference on Automation Science and Engineering
metering of cloud services,” in 6th Annual IEEE International
(IEEE CASE 2010), Toronto, Canada, Aug. 2010, pp. 293–300,
Systems Conference (IEEE SysCon 2012), Vancouver, BC, Canada,
doi:10.1109/COASE.2010.5584153.
Mar. 2012, doi:10.1109/SysCon.2012.6189462.
[8] Q. Chen, P. Grosso, K. van der Veldt, C. de Laat, R. Hofman, [19] A. Qureshi, R. Weber, H. Balakrishnan, J. Guttag, and B. Maggs,
and H. Bal, “Profiling energy consumption of vms for green “Cutting the electric bill for internet-scale systems,” SIGCOMM
cloud computing,” in 2011 IEEE Ninth International Conference on Comput. Commun. Rev., vol. 39, no. 4, pp. 123–134, Aug. 2009, doi:
Dependable, Autonomic and Secure Computing (DASC), Dec. 2011, 10.1145/1594977.1592584.
pp. 768–775, doi: 10.1109/DASC.2011.131.
[20] T. Heath, B. Diniz, E. V. Carrera, W. Meira, Jr., and R. Bianchini,
[9] A. Kansal, F. Zhao, J. Liu, N. Kothari, and A. A. Bhattacharya, “Energy conservation in heterogeneous server clusters,” in
“Virtual machine power metering and provisioning,” in Proceedings of the Tenth ACM SIGPLAN Symposium on Principles
Proceedings of the First ACM Symposium on Cloud Computing, ser. and Practice of Parallel Programming, ser. PPoPP ’05. ACM, 2005,
SoCC ’10. ACM, 2010, pp. 39–50, doi: 10.1145/1807128.1807136. pp. 186–195, doi: 10.1145/1065944.1065969.
[10] S. Bruno, S. Lamonaca, M. La Scala, G. Rotondo, and [21] M. Davison, C. Anderson, B. Marcus, and K. Anderson,
U. Stecchi, “Load control through smart-metering on distribution “Development of a hybrid model for electrical power spot prices,”
networks,” in 2009 IEEE PowerTech, July 2009, pp. 1–8, doi: IEEE Trans. Power Syst., vol. 17, no. 2, pp. 257–264, May 2002.
10.1109/PTC.2009.5282256. [22] F. Alvarado and R. Rajaraman, “Understanding price volatility in
[11] R. van Gerwen, S. Jaarsma, and R. Wilhite, “Smart metering,” electricity markets,” in 2000. Proceedings of the 33rd Annual Hawaii
[Online], Available at http://www.leonardo-energy.org/webfm International Conference on System Sciences, Jan. 2000, p. 5 pp.
send/435. [23] N. Domingues, “Influence of the renewable energy in the
[12] “Global Data Center Energy Demand Forecasting,” simulation of liberalised markets,” in Computational Sciences and
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
IEEE TRANSACTIONS ON SERVICES COMPUTING
14
Optimization, 2009. CSO 2009. International Joint Conference on, [45] D. C. Snowdon, S. M. Petters, and G. Heiser, “Accurate on-line
vol. 2, April 2009, pp. 480 –484. prediction of processor and memory energy usage under voltage
[24] Q. Zhao, P. Wang, Y. Ding, and L. Kumar Goel, “Impacts of scaling,” in Proceedings of the 7th ACM & IEEE International
solar power penetration on nodal prices and nodal reliability,” in Conference on Embedded Software, ser. EMSOFT ’07. ACM, 2007,
IPEC, 2010 Conference Proceedings, Oct. 2010, pp. 1134–1139, doi: pp. 84–93, doi: 10.1145/1289927.1289945.
10.1109/IPECON.2010.5696993. [46] W. Bircher and L. John, “Complete system power estimation using
[25] C. Lavania, S. Rao, and E. Subrahmanian, “Reducing processor performance events,” IEEE Trans. Comput., vol. 61, no. 4,
variation in solar energy supply through frequency domain pp. 563–577, April 2012, doi: 10.1109/TC.2011.47.
analysis,” IEEE Syst. J., vol. 6, no. 2, pp. 196–204, Jun. 2012, [47] A. Kivity, Y. Kamay, D. Laor, U. Lublin, and A. Liguori, “kvm:
doi:10.1109/JSYST.2011.2162796. the linux virtual machine monitor,” Online, Tech. Rep., 2007,
[26] W. Katzenstein, E. Fertig, and J. Apt, “The variability of available at https://www.kernel.org/doc/mirror/ols2007v1.pdf#
interconnected wind plants,” Energy Policy, vol. 38, no. 8, pp. page=225.
4400–4410, 2010, doi: 10.1016/j.enpol.2010.03.069. [48] R. Buyya, D. Abramson, and J. Giddy, “A case for economy grid
[27] D. Hart, “Using ami to realize the smart grid,” in 2008 IEEE Power architecture for service oriented grid computing,” International
and Energy Society General Meeting - Conversion and Delivery of Parallel and Distributed Processing Symposium, vol. 2, 2001.
Electrical Energy in the 21st Century, July 2008, pp. 1–2. [49] L. McKnight and J. Boroumand, “Pricing internet services:
[28] “Assessment of demand response and advanced metering,” proposed improvements,” Computer, vol. 33, no. 3, pp. 108–109,
Federal Energy Regulatory Committee, Tech. Rep. March 2000.
[Online]. Available: http://www.ferc.gov/legal/staff-reports/ [50] “Amazon pricing,” [Online], Available at http://aws.amazon.
2010-dr-report.pdf com/ec2/pricing/.
[29] H.-P. Siderius and A. Dijkstra, “Smart metering for households: [51] “Rackspace pricing,” [Online], Available at http://www.
Cost and benefits for the Netherlands,” EEDAL, 2006. rackspace.com/cloud/cloud hosting products/servers/
[30] J. W. Ross, M. R. Vitale, and C. M. Beath, “The untapped potential pricing/.
of it chargeback,” MIS Quarterly, vol. 23, no. 2, pp. 215–237, June [52] “Hypervisor statistics counter definition,” [Online], Available
1999. at http://msdn.microsoft.com/en-us/library/windows/
[31] R. Iyer, R. Illikkal, L. Zhao, D. Newell, and J. Moses, “Virtual hardware/ff541603\%28v=vs.85\%29.aspx.
platform architectures for resource metering in datacenters,” [53] “Hyperic sigar api,” [Online], Available at http://www.hyperic.
SIGMETRICS Perform. Eval. Rev., vol. 37, pp. 89–90, October 2009. com/products/sigar.
[32] “Amazon spot instance,” [Online], Available at http://aws. [54] “Hyperic opensource monitoring,” [Online], Available at http://
amazon.com/ec2/spot-instances/. www.hyperic.com/products/open-source-systems-monitoring.
[55] “Nagios,” [Online], Available at http://www.nagios.org/.
[33] T. Singh and P. K. Vara, “Smart metering the clouds,” in 18th
[56] “Open eucalyptus,” [Online], Available at http://open.
IEEE International Workshops on Enabling Technologies: Infrastructure
eucalyptus.com/.
for Collaborative Enterprises (WETICE 2009), 2009, pp. 66–71,
[57] “Ubuntu cloud images,” [Online], Available at http:
doi:10.1109/WETICE.2009.49.
//uec-images.ubuntu.com/.
[34] A. H. Mohsenian-Rad and A. Leon-Garcia, “Optimal residential
[58] “Apache HTTP server,” [Online], Available at http://httpd.
load control with price prediction in real-time electricity pricing
apache.org/.
environments,” Smart Grid, IEEE Transactions on, vol. 1, no. 2, pp.
[59] “MySQL open source database server,” [Online], Available at
120–133, September 2010, doi: 10.1109/TSG.2010.2055903.
http://www.mysql.com/.
[35] J. Rao and C.-Z. Xu, “Online capacity identification of [60] “Pjm interconnection llc,” [Online], Available at https://edata.
multitier websites using hardware performance counters,” IEEE pjm.com/eData/index.html.
Trans. Parallel Distrib. Syst., vol. 22, pp. 426–438, 2011, doi:
10.1109/TPDS.2010.92.
[36] C. Lefurgy, K. Rajamani, F. Rawson, W. Felter, M. Kistler, and T. W.
Keller, “Energy management for commercial servers,” Computer,
vol. 36, pp. 39–48, 2003, doi: 10.1109/MC.2003.1250880.
[37] R. Bianchini and R. Rajamony, “Power and energy management
for server systems,” Computer, vol. 37, pp. 68–74, 2004, doi: Akshay Narayan completed his MTech in
10.1109/MC.2004.217. information technology from IIIT-Bangalore
[38] P. Ranganathan, P. Leech, D. Irwin, and J. Chase, “Ensemble-level and currently pursuing his Ph.D in School of
power management for dense blade servers,” SIGARCH Comput. Computing, National University of Singapore.
PLACE He is currently working on sequential decision
Archit. News, vol. 34, no. 2, pp. 66–77, May 2006, doi: PHOTO
10.1145/1150019.1136492. making. His research interests include artificial
HERE intelligence and its application in cloud
[39] Q. Deng, D. Meisner, L. Ramos, T. F. Wenisch, and R. Bianchini,
“Memscale: active low-power modes for main memory,” computing.
SIGARCH Comput. Archit. News, vol. 39, no. 1, pp. 225–238, Mar.
2011, doi: 10.1145/1961295.1950392.
[40] J. Choi, S. Govindan, J. Jeong, B. Urgaonkar, and
A. Sivasubramaniam, “Power consumption prediction and
power-aware packing in consolidated environments,” IEEE Trans.
Comput., vol. 59, pp. 1640–1654, 2010, doi: 10.1109/TC.2010.91.
[41] J.-W. Jang, M. Jeon, H.-S. Kim, H. Jo, J.-S. Kim, and
S. Maeng, “Energy reduction in consolidated servers through Shrisha Rao received his Ph.D. in computer
memory-aware virtual machine scheduling,” IEEE Trans. Comput., science from the University of Iowa, and before
vol. 60, pp. 552–564, 2011, doi: 10.1109/TC.2010.82. that his M.S. in logic and computation from
[42] D. Economou, S. Rivoire, and C. Kozyrakis, “Full-system power Carnegie Mellon University. He is an associate
PLACE professor at IIIT-Bangalore, a graduate school of
analysis and modeling for server environments,” in In Workshop
PHOTO information technology in Bangalore, India. His
on Modeling Benchmarking and Simulation (MOBS), 2006.
HERE research interests are in distributed computing,
[43] J. C. McCullough, Y. Agarwal, J. Chandrashekar, S. Kuppuswamy,
A. C. Snoeren, and R. K. Gupta, “Evaluating the effectiveness of specifically algorithms and approaches for
model-based power characterization,” in Proceedings of the 2011 concurrent and distributed systems, and include
USENIX conference on USENIX annual technical conference, ser. solar energy and microgrids, cloud computing,
USENIXATC’11. USENIX Association, 2011, pp. 12–12. energy-aware computing (”green IT”), and
[44] B. Krishnan, H. Amur, A. Gavrilovska, and K. Schwan, demand side resource management. Dr. Rao is a member of the IEEE
“Vm power metering: Feasibility and challenges,” SIGMETRICS Computer Society, the ACM, the American Mathematical Society, and
Perform. Eval. Rev., vol. 38, no. 3, pp. 56–60, Jan. 2011, doi: the Computer Society of India.
10.1145/1925019.1925031.