Procurement Auctions to Maximize Players’ Utility in Cloud Markets
Paolo Bonacquisto, Giuseppe Di Modica, Giuseppe Petralia and Orazio Tomarchio
Department of Electrical, Electronic and Computer Engineering, University of Catania, Catania, Italy
Keywords:
Cloud Market, Procurement Auction, Bidding Strategy, Cloud Simulations.
Abstract:
Cloud computing technology has definitely reached a high level of maturity. This is witnessed not just by
the interest of the academic community around it, but also by the wide adoption in commercial scenarios.
Today many big IT players are making huge profits from leasing their computing resources “on-demand”, i.e.,
letting customers ask for the amount of resources they need, and charging them a fixed price for each hour
of consumption. Recently, studies authored by economists have criticized the fixed-price applied to cloud
resources, claiming that a better pricing model can be devised which may increase profit for both the vendor
and the consumer. In this paper we investigate how to apply the mechanism of procurement auction to the
problem, faced by providers, of allocating unused resources. In particular, we focus on the strategies providers
may adopt in the context of procurement auctions to maximize the utilization of their data centers. We devised
a strategy, which dynamically adapts to changes in the auction context, and which providers may adequately
tune to accommodate their business needs. Further, overbooking of resources is also considered as an optional
strategy providers may decide to adopt to increase their revenue. Simulations conducted on a testbed showed
that the proposed approach is viable.
1 INTRODUCTION
Cloud computing aims to provide computing re-
sources to customers like public utilities such as
water and electricity (Buyya et al., 2008). In an
Infrastructure-as-a-Service (IaaS) cloud environment,
physical resources are packaged into distinct types
of virtual machines (VMs) and offered to customers.
A cloud customer, on the other hand, will purchase
VMs to run his applications, by looking for specific
resource requirements in terms of CPU, memory and
disk. Given the finite capacity for each type of re-
sources in each data center, a fundamental problem
faced by IaaS provider is how to select the price and
allocate resources for each type of VM services in or-
der to best match the interests of the customers while
maximizing his revenue. This issue is further com-
plicated by the fact that, differently from traditional
utility markets, cloud demand is strongly time vary-
ing and often burstly.
The resource allocation and trading mechanisms
used by the current cloud computing systems are in-
efficient and inflexible due to the flat rate pricing
model adopted. We argue that a fixed price-based re-
source allocation currently in use in cloud comput-
ing systems do not provide an efficient allocation of
resources and do not maximize the revenue of the
cloud providers. In a previous work (Di Modica
et al., 2013), we already showed that a better alter-
native would be to use auction-based resource allo-
cation mechanisms. In this paper we address issues
related to the bidding strategies adopted by providers
of computing resources in the context of procurement
auctions. We try to analyze all the factors that mainly
impact the strategic choices of providers in the acqui-
sition of the goods allocated through auctions. The
purpose of this work is not to devise an optimal bid-
ding strategy, but rather, to prove that any strategy
will have its objective guaranteed by the procurement
mechanism. We also devised a tentative provider’s
strategy which adapts its aggressiveness to the earlier
mentioned factors. In the addressed market scenario,
we stress that our attention is devoted to the optimiza-
tion of the utilization rate of providers’ data centers
and the utility of providers.
The remainder of the paper is structured as fol-
lows. Section 2 makes a review of the literature and
gives some rationale of the work. Section 3 intro-
duces the proposed idea and delves into technical de-
tails about procurement auctions. In Section 4 simu-
lation results are presented and discussed. Finally, the
work is concluded in Section 5.
38
Bonacquisto P., Di Modica G., Petralia G. and Tomarchio O..
Procurement Auctions to Maximize Players’ Utility in Cloud Markets.
DOI: 10.5220/0004854700380049
In Proceedings of the 4th International Conference on Cloud Computing and Services Science (CLOSER-2014), pages 38-49
ISBN: 978-989-758-019-2
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
2 RELATED WORK
All main commercial IaaS providers “comply” with
the On-Demand approach, and offer to charge cus-
tomers for the actual time frame during which the
resource is actually utilized
1
. They ask users pay
a fixed price for computing capacity by the hour.
The only chance for the customer to get discounts
on the price-per-hour is to opt on the reservation or
the flat rate charging models. According to these op-
tions, customers get a discount on the price but are
committed to longer periods of lease (from a month
to a year). Among commercial providers, the only
one that successfully proposed an approach alterna-
tive to the fixed-price is Amazon with its Spot In-
stance model
2
. This model enables the customer to
bid for what they call unused computing capacity.
Virtual machines are charged the Spot Price, which
is set by Amazon and fluctuates periodically depend-
ing on the supply/demand rate for computing capac-
ity. According to this model, on the one hand the
provider gives the customer the possibility to acquire
computing resources at a lower price then the stan-
dard; on the other one, whenever the resources’ de-
mand increases, the provider reserves the right to pre-
empt those resources and give them to better bidders.
This model represents the very first attempt to build
up a virtual market of computing resources regulated
by market prices, i.e., prices which dynamically fluc-
tuate according to offer and demand. In spite of this,
the model is still unclear (the formula of price fluc-
tuation is not known) and is not proved to be re-
sistant to potential malicious behaviors of customers
(dishonest customers can abuse the system and ob-
tain short-term advantages by bidding large maximum
price bid while being charged only at the lower spot
price (Wang et al., 2012)). Furthermore in (Agmon
Ben-Yehuda et al., 2011) authors prove that the Ama-
zon’s Spot Price is not market driven, rather is typi-
cally generated as a random value near to the hidden
reserve price within a tight price interval. The con-
sideration stemming from this observation is that a
provider, being an interested party, may not be a guar-
antee for the correctness of the price determination.
Instead, a third party broker should be in charge of
calling out prices in auction-based contexts. A quick
review of the recent literature proves that researchers
are very much concerned with the application of auc-
tion mechanisms to the problem of the allocation
(read “sale”) of computing resources. In (Risch et al.,
1
http://aws.amazon.com/ec2/
http://www.microsoft.com/windowsazure/
http://www.rackspace.com/
2
http://aws.amazon.com/ec2/spot-instances/
2009) authors propose a marketplace of computing
resources where prices are determined using an ex-
change market. In (Chard and Bubendorfer, ) authors
discuss several strategies that cloud providers should
adopt in order to reach high performance and to over-
come most of criticisms of auctions like high over-
heads and high latency using techniques like over-
booking and Flexible Advanced Reservations. They
propose several bidding functions but each one takes
into account only one parameter among those moni-
tored by a cloud provider. Our work is different as
we try to guarantee the provider’s utility maximiza-
tion through an adaptive strategy based on several pa-
rameters referring to the actual condition of resources
allocation, intentionally weighted by the provider in
order to address a specific target.
For the majority of researchers, combinatorial
auctions are the most appropriate sale mechanism for
allocating virtual machines in the cloud. In combi-
natorial auctions the participants bid for bundles of
items rather than individual items (Cramton et al.,
2005). This mechanism seems to perfectly fit the
Cloud context, as customers usually need to acquire
not just one resource but a bunch of resources (e.g.,
one for hosting the database server, one for the appli-
cation server and one for the web server). In (Wang
et al., 2012) authors propose a suite of computation-
ally efficient and truthful auction-style pricing mech-
anisms, which enable customers to fairly compete for
resources and cloud providers to increase their overall
revenue. (Zaman and Grosu, 2013) proposes a com-
binatorial auction-based protocol for resource alloca-
tion in grids. They considered a model where dif-
ferent grid providers can provide different types of
computing resources. Buyya et al. (Buyya et al.,
2010) propose an infrastructure for auction-based re-
source allocation across multiple cloud systems. In
(Vinu Prasad et al., 2012) authors address the sce-
nario of multiple resource procurement in the realm of
cloud computing. In the observed context, they pre-
process the user requests, analyze the auction and de-
clare a set of vendors bidding for the auction as win-
ners based on the Combinatorial Auction Branch on
Bids (CABOB) model.
The discussed works mainly focus on solving the
problem of optimal sale of resources in combinato-
rial auctions, which is known to be NP-hard. The
work we propose, instead of defining yet another sub-
optimal allocation algorithm, takes a different direc-
tion. From a strictly technical point of view, one of
IaaS providers’ main issue is to adopt an efficient allo-
cation scheme which allows them to map customers’
requests to virtual machines (or virtual resources, in
general) in an efficient way. Often, when a new re-
ProcurementAuctionstoMaximizePlayers'UtilityinCloudMarkets
39
quest must be served there are many management op-
erations that need to be carried out along. Let us sup-
pose that an IaaS provider has a number of hosts, and
a new request R demanding some computing power
has arrived. According to both the actual hosts’ oc-
cupancy rate and the adopted allocation scheme many
different actions may be taken. For instance, comput-
ing capacity may be allocated in the host where the
computing availability best (or worst) fits the comput-
ing request. Or, computing capacity may be allocated
in an unloaded host running on a stand-by state. Or
again, a running virtual machine allocated in host H
1
may be migrated to host H
2
in order to make room
for R (whose size fits better in H
1
than in H
2
). Those
cited here are only a few of the many examples of
management issues that providers must face with.
The profit of a provider strongly depends on its
capability of keeping the hosts’ average occupancy
rate as high as possible. For their nature, comput-
ing resources can be regarded as perishable goods that
need to be sold within a certain time frame otherwise
they get wasted. Not selling a virtual machine in a
given slot time means a profit loss for the provider,
who anyway is spending money to keep the physi-
cal machines up and running. We then look at the
trade of computing resources from a new perspective,
in which providers, in the aim of maximizing their
data center’s occupancy rate, may be willing to attract
customers by lowering the offer price. On their turn,
customers may get what they need, at the time they
need it, at a price which is lower than the standard
price at which they usually buy. We advocate that
the market model best fitting this perspective is the
one which guarantees the sale of computing resources
through procurement auctions. Procurement auc-
tions (Klemperer, 1999) (also called reverse auctions)
reverse the roles of sellers and buyers, in the sense
that the bidders are those who have interest in selling
a good (the providers), and therefore the competition
for acquiring the right-to-sell the good is run among
providers.
The sale of computing resources through procure-
ment auctions will work as follows. The market gath-
ers computing demands from any customers. For each
computing request a procurement auction is publicly
called out. Providers can search the market in order
to identify the specific request(s) which best fits their
need in that particular moment (e.g., a request for a
virtual machine of a specific size that would “fill” a
given host’s capacity) and participate in the respec-
tive call(s). A call will be won by the provider offer-
ing to serve the demand at the lowest price. On the
provider’s side, this mechanism is profitable since the
huge number of computing demands gathered by the
market increases the chances of the provider to find
the one(s) satisfying their needs. In their turn, many
customers will be attracted by the possibility to get
what they ask at a price that is lower than the standard,
therefore will be stimulated to push their requests to
the procurement-based market.
3 PROCUREMENT AUCTION
MARKET
The purpose of this work is not to convince providers
to abandon the direct-sell mechanism in favor of the
procurement-based market. Providers have their reg-
ular customers, who issue requests which most of
the times have a well known timing. For this kind
of requests the most appropriate model is the direct-
sell/fixed-price, in that it provides guarantees for both
the provider and the customer. What we propose is
the adoption of an alternative, dynamic pricing model
for selling what is usually referred to as the unused
capacity, i.e., the residual capacity that, on average,
the provider is not able to sell through direct-sell.
Let us define the utilization rate U(t) as the frac-
tion of the overall unused capacity committed to serve
customers’ requests at time t . The lower U, the
higher the profit loss for the provider. In the aim
of maximizing the utilization rate (minimizing the
residual capacity) providers need to adopt new selling
strategies. The simplest strategy could just be lower-
ing the price per computing unit. Amazon currently
leases its unused capacity by adopting an auction-
inspired price strategy that let the customers acquire
resources for a price which is lower than the standard.
We argue that providers, to avoid “wasting” comput-
ing capacity, are willing to give up a portion of profit
per computing unit (same as it happens for sale of per-
ishable goods). As far as we know, the Amazon’s is
the only example of dynamic price strategy that is al-
ternative to the fixed-price. If on the one hand it is true
that customers benefit from low prices, on the other
one the proprietary mechanism by which the virtual
machines’ price fluctuates has never been disclosed.
The computing capacity’s actual supply/demand rate
is not shared to customers, nor the price policy has
ever been released.
In this paper we propose the design of a market of
computing capacity (Figure 1), to which any provider
is admitted, and where computing resources can be
sold through auction-based allocation schemes. The
perspective is that of procurement auctions, where an
initial price is called out and bidders iteratively have
to call lower prices to win. The market mechanism
is the following. Customers communicate their com-
CLOSER2014-4thInternationalConferenceonCloudComputingandServicesScience
40
Figure 1: The Cloud Marketplace scenario.
puting demand to the market. A broker will take care
of demands. For each specific demand, the broker
(auctioneer) will run a public auction to which any
provider (bidder) can participate and compete for “ac-
quiring” the demand. The winning provider (who of-
fered the lowest price) will eventually have to serve
the demand. Being the auctions open to the participa-
tion of multiple providers, the competition is granted.
Providers will have to fight to gain the right-to-serve
the demand. Bidding strategies enforced by providers
can range from the most conservative to the most ag-
gressive. The determination of the final price is driven
only by the evaluation that each provider has on the
goods to acquire (i.e., the customer’s demand to be
served). Customers will get their demand served at
the lowest price. Further, they will have no more the
burden to search for providers, as providers are gath-
ered in the market. As for the broker, its income will
be evaluated in a way proportional to both the number
of transactions (auctions) it will be able to carry out
and the prices formed in the transactions.
3.1 Customer’s Demand
Customers need to acquire cloud resources to sat-
isfy their computing needs. Following the cloud
paradigm, the computing power of a physical ma-
chine can be adequately shaped to satisfy the most
demanding customer need, in terms of CPU cores,
CPU cycles, Mips and main memory size. In practice,
all cloud providers offer a restricted number of vir-
tual machines (VM) configurations (instance types),
which go from the very minimalist to the most pow-
erful one. In this paper we will refer to a classification
of VM instance types which recalls that of Amazon.
Though discrepancies exist between the classification
being presented and those adopted by other providers,
schemes that map the respective items can easily be
derived (but are out of the scope). In the rest of the
paper, we will then assume that customers’ demand
will address the following VM instance types:
General Purpose.
M1.small - 32/64-bit architecture, 1 vCPU,
1 ECU, 1.7GB RAM, 160GB Storage, Low
Bandwidth;
M1.medium - 32/64-bit architecture, 1 vCPU,
2 ECU, 3.75GB RAM, 410GB Storage, Mod-
erate Bandwidth;
M1.large - 64-bit architecture, 2 vCPU, 4 ECU,
7.5GB RAM, 820GB Storage, Moderate Band-
width;
M1.xlarge - 64-bit architecture, 4 vCPU, 8
ECU, 15GB RAM, 1.6TB Storage, High Band-
width;
M3.xlarge - 64-bit architecture, 4 vCPU, 13
ECU, 15GB RAM, 0 Storage, Moderate Band-
width;
Compute Optimized.
C1.medium - 32/64-bit architecture, 2 vCPU, 5
ECU, 1.7GB RAM, 350GB Storage, Moderate
Bandwidth;
C1.xlarge - 64-bit architecture, 8 vCPU, 20
ECU, 7GB RAM, 1.6TB Storage, High Band-
width;
Memory Optimized.
M2.xlarge - 64-bit architecture, 2 vCPU, 6.5
ECU, 17.1GB RAM, 420GB Storage, Moder-
ate Bandwidth.
We also assume that customers may demand for
combinations of the above listed resources. For in-
stance, a customer may issue a request for three
M2.xlarge instances and four M1.large instances.
Along with the (combination of) resources needed,
the customer must communicate the time at which the
resource(s) need to be accessed (start-provision time)
and the time at which the resource provision must ter-
minate (stop-provision time).
3.2 Auction Types
Auctions are the tool that the broker will make use
of in order to allocate providers’ computing capacity
to customers’ demand. An auction can be run by the
broker as soon as a new demand has arrived. For a
given demand, the broker can choose from some auc-
tion types to run. The choice of the best suiting auc-
tion type is driven by several factors, which are object
of observation in this paper. Also, it is up to the broker
the setting of the initial price of the auction.
As outlined before, the typology of auction that
best fits the depicted context is the procurement auc-
tion. In a procurement auction the buyer (customer,
in our case) advertises a need (demand for computing
capacity) and sellers (providers) compete to provide
the service (virtual machines to satisfy the demand).
ProcurementAuctionstoMaximizePlayers'UtilityinCloudMarkets
41
This leads to decreasing prices being offered during
the period of the procurement auction.
In this paper we will focus on three different types
of procurement auctions. The common part of the
three auction mechanisms is the auction preparation,
which provides that upon the arrival of a demand, the
broker issues a public “call for proposal” (CFP) to
invite providers. The CFP shall specify a minimum
set of auction parameters including the start-provision
time, the stop-provision time, the initial price (from
which discount bids are expected), the bidding rules
(who can bid and when, restrictions on bids) and the
clearing policy (when to “terminate” the auction, who
gets what, at what price). After having collected the
willingness of providers to participate in the auction,
the preparation phase ends up and the bargain starts
according to what is specified in the bidding rules
and the clearing policy. What characterizes one auc-
tion mechanism from another one is the information
specified in the bidding rules and the clearing policy.
For our purpose we will address the following mech-
anisms:
English Reverse (ER)(Parsons et al., 2011). The
ER is a multi-round auction. The CFP specifies
the initial price from which discounting bids (of-
fers) are expected. The participating bidders can
post their offers. Discounting offers are called out,
so that every bidders is always aware of the refer-
ence price for which further discounts are to be
proposed. If no offer arrives within a time-frame
(publicly set in the CFP), the good will be as-
signed to the last best (i.e., the lowest priced) of-
fer. This type of auction allows bidders gather in-
formation of each other’s evaluation of the good.
First Price Sealed Bid (FPSB)(Parsons et al.,
2011). The FPSB is a single-round auction, i.e.,
all bidders have the chance to bid just once, before
the auction being cleared. When bidders receive
the CFP, they check the initial price and decide to
either bid or not to bid. After all participants have
posted their bid, the broker clears the auction and
allocates the “demand” to the bidder who has val-
ued it the most (in the case of a reverse auction,
the least). The peculiarity of this auction is that
bidders are not aware of each other’s offer, as only
the winning bid will be broadcast at the end of the
auction.
Second Price Sealed Bid (SPSB). Like the FPSB,
it is a single round auction. But with the differ-
ence that the one who wins the auction (i.e., who
offers the highest price) will pay the second high-
est bid price. This mechanism, applied to a re-
verse auction, aims at improving the provider’s
utility (refer to the next section for the definition
of the provider’s utility). In fact, according to
this mechanism each provider may keep its good’s
evaluation secret, and if they will win the good
they will be acknowledged a price which is higher
than their own bid, thus increasing the overall util-
ity.
3.3 Provider’s Strategy
One of the objective of this work is the study of an
adaptive (i.e., dynamic) strategy for the providers that
participate in procurement auctions. By strategy we
mean a set of rules producing the decisions a provider
must take to maximize their own business objective.
Basically, a strategy shall drive the provider in choos-
ing the right actions to be undertaken when competing
for the acquisition of a good (e.g., whether to partici-
pate in a given auction, to bid in a given round, not to
bid, which price to offer). In the strategy design, the
first step was to outline the main factors that may im-
pact such choices. Secondly, we tried to devise a dy-
namic strategy which accounts for the just mentioned
factors and smoothly adapt their fluctuations. Finally,
we set up and configured a test environment to ana-
lyze the results produced by the strategy.
According to the literature, the behavior of an auc-
tion’s participant is mainly driven by the information
the participant has on the value of the good being
sold (Klemperer, 1999). In respect to this informa-
tion, two basic auction models are possible: 1) the
private-value model, where each bidder has an esti-
mate of the good for sale, and that estimate is private
and unaffected by others’ estimates, and 2) the pure
common-value model, where the actual value of the
object is the same for everyone, but bidders have dif-
ferent private information on how much that value ac-
tually is. Combined models can also be derived from
the cited ones.
If we better analyze the context of cloud auctions,
a computing resource can be seen as a good whose ac-
tual value (price) is common to all providers. In fact,
though for computing resources we can not yet speak
of conventional “market prices”, all providers in their
regular sales adopt well known, leveled prices. We
can then conclude the actual values of such kind re-
sources are somewhat common to providers. In the
context of a procurement auction of computing re-
sources, the estimate E
pi
of the i-th provider for a
given good may differ from the the estimate E
p j
of
the j-th provider according to the diverse needs each
provider may have in pursuing their own business ob-
jective.
Primary objective of a provider is to maximize
what is referred to as Utility. Given a resource to
CLOSER2014-4thInternationalConferenceonCloudComputingandServicesScience
42
be allocated through an auction sale, the provider’s
Utility for that request is defined as the difference be-
tween the winning bid price and the evaluation that
the provider gives to the resource. (Wang et al., 2012).
Of course, the provider aims at maximizing the aver-
age utility for the resources they compete for. Recall-
ing the considerations made earlier, in the context of
an auction sale of spare resources this objective can
be pursued: a) by keeping the data center’s utiliza-
tion as high as possible; b) by bidding prices higher
than the personal evaluation (which we will refer to as
lower bound) and c) by choosing the most profitable
combination of customers’ request to serve.
We identified a non-exaustive list of factors which
may strongly influence the strategy of a provider in a
procurement auction.
The duration of the customer task (demand) to be
served (L). The longer the task, the higher the
profitability for the provider, since the required
capacity will be committed for a longer time. A
provider, then, might prefer to participate in auc-
tions where long tasks are traded.
The type of VM instance required to serve the cus-
tomer task (T ). Of course, the profitability of a
task is directly proportional to the task’s require-
ments in terms of amount of computing capacity
per hour, so providers may be motivated in point-
ing on auctions calling for a higher capacity/hour.
But depending on the actual utilization level of
both each single host and the whole data center,
it might not be possible to serve further tasks re-
quiring high capacity VMs.
The gap between the potential revenue obtain-
able from serving the task the standard way (i.e.,
through the fixed-price market) and that obtain-
able by serving the task at the price called by the
auctioneer (G
r
). The revenue for serving a task is
given by the L times the price (P) of the resource
that will serve the task. This factor strongly de-
pends on the provider’s enforced revenue policy.
A provider pointing on auctions to sell their un-
used capacity might accept a much lower revenue
(bidding a lower price) in the case that expenses
are already covered. Conversely, the provider
might not be willing to excessively lower the price
in the case that expenses are not yet covered.
The utilization of the particular physical machine
that is going to serve the customer’s request. The
marginal revenue, in fact, is affected by the uti-
lization level of a host: if a host is already running
and serving other tasks, adding more tasks to that
host “costs” less than activating a new host.
Finally, some considerations need to be made
about the lower bound. Each strategy must envision
an “exit condition”, which represents the condition
that, when verified, forces the exit of the provider
from the auction. When the provider decides to par-
ticipate in an auction, they will have to set the lower
bound price, which represents the maximum discount
that the provider is willing to offer for the good be-
ing traded in that auction. Of course, this parame-
ter only makes sense in multi-rounds auctions, as in
single round auctions actually exit is imposed by the
mechanism itself at the end of the first round. The
lower bound parameter actually represents the eval-
uation of the provider for a given good (customer’s
request). It incorporates all provider’s consideration
regarding the costs for executing a VM, managing a
VM’s life cycle and supporting the customer.
The objective of a strategy is to suggest the
provider the price to call for the next bid. In calling
a price, a strategy may be more or less “aggressive”,
i.e., may propose higher or lower discounts. We dis-
cuss two different strategies. One is driven by ran-
domness (aggressiveness is randomly chosen auction
by auction, round by round). The other is adaptive,
in the sense that is able to adapt the aggressiveness
according to the above listed factors. For this kind
of strategy, the aggressiveness can be tuned by ade-
quately weighting the factors.
Recalling a formula presented in (McAfee, 1987),
the adaptive strategy will suggest the next bid as:
bid =
n 1
n (1 α)
lastWinningBid (1)
where n is the number of bidders participating in
the auction and lastWinningBid is the price of the bid
that won the last round. In case of single-round auc-
tions and in the case of first rounds of multi-round
auctions, lastWinningBid will be the auction’s start-
ing price. The parameter α is calculated as follows:
α = w
1
U (t) + w
2
L
L
max
+ w
3
P
a
P
f
+ w
4
T
vm
T
max
+ w
5
H(t) (2)
As we can see in the equation 2, α depends on:
U(t), the current utilization rate of the pool of
spare resources; the less U (t), the higher α, so
the evaluated bid price will decrease (in a reverse
auction, lowering the bid price means pointing to
gain the good). As expected, the aggressiveness
of a strategy increases with the reduction of the
utilization rate.
L
L
max
, the ratio between the time period for which
the computing resource is requested and the max-
imum time period for which a resource can be
ProcurementAuctionstoMaximizePlayers'UtilityinCloudMarkets
43
requested
3
. The ratio will increase for requests
with longer execution time. The provider will
be more aggressive in auctions where longer cus-
tomer tasks are negotiated, as those ensure a
higher utilization of the data center and, as a con-
sequence, higher revenues.
P
a
P
f
the ratio between the resource’s starting price
in the auction and the corresponding price in the
standard fixed-price market. The provider’s ag-
gressiveness will be higher when the price at the
start of a round is closer to the reference price
(price at which resources are traded in regular
markets, or, direct-sell price). The more the round
price decreases, the lesser the provider’s aggres-
siveness.
T
vm
T
max
the ratio between the computing power of the
resource being traded in the auction and the com-
puting power of the highest resource. This fac-
tor increases the provider’s aggressiveness in the
case of customer tasks demanding high comput-
ing power. The higher the requested computing
power, the higher the task’s initial price. Further,
a highly demanding task requires a bigger capac-
ity on the data center, thus increasing the overall
utilization rate.
H(t), the current utilization of the host on which
the customer task to serve will be scheduled. This
factor increases the α parameter and, therefore,
increases the provider’s aggressiveness. Recall-
ing a previous consideration, the provider is more
conservative in their strategy if for serving a task
a new physical machine has to be activated.
Each parameter is weighted by a factor
(w
1
,w
2
,w
3
,w
4
,w
5
), for which the following con-
straint applies:
5
i=1
w
i
= 1 (3)
Different combinations of weights lead to differ-
ent strategies. Finally, in the adaptive strategy the
lower bound price will depend on α according to the
following equation:
Lb = P
f
(1 discount) (4)
where discount is
discount = (0.5 α) ± rand 0.03 (5)
and P
f
is the price of the resource advertised in the
standard fixed-price market. The maximum discount
3
In real situations the time period for which a resource
can be requested has no bound; in our simulation we will
take into account tasks lasting no longer than 24 hours
on the fixed price is evaluated as the 50% of α; the
higher alpha, the lower the bound. A variability of
3% was also introduced to model a differentiation
among providers, which reflects their respective per-
sonal evaluations.
3.4 Resource Overbooking
The auction mechanism causes a waste of computing
resources at the provider’s end. A provider may par-
ticipate in many auctions (say m) at the same time.
For each auction, no matter they win or lose, the
provider will have to reserve a pool of resources to ac-
commodate the customer’s request for which they are
competing. The number of auctions every provider
will participate depends on the instant capacity of
their free computing resources. In general, provider
will win n auctions, being n <= m, thus, for the du-
ration of all m auctions there may be a waste of re-
sources proportional to the number of lost auctions
(m n). To overcome this limitation, the provider
may decide to participate in more auctions and com-
pete for customers’requests which they are not poten-
tially able to meet. This mechanism, also known as
resource overbooking, contributes to decrease the re-
source waste on the one hand, but on the other may
bring to situations where the provider runs out of
computing resources and may not honor one or more
contracts signed at the time they won the auctions. In
this cases, the risk compensation principle is applied
(Phillips, 2005), and the provider will incur penalties
which are proportional to their actual bid.
In order to implement such mechanism in our mar-
ket, we let the provider count on an amount of vir-
tual computing capacity (namely, overbooking capac-
ity) which is set to 20% of their real capacity. The
provider is then able to participate in m + o auctions,
where o is proportional to the overbooking capacity.
This way the number of won auctions will increase,
and the provider’s utilization rate will get closer to
1. In the case the provider won more auctions than
those they are actually able to serve, a penalty is due.
When an auction appoints as winner a provider who
is not eventually able to honor a request, the second
best bidding provider is chosen. If, again, the latter is
not able to serve the request, the third best is chosen,
and so on. In this chain, all providers are subjected to
penalties. The penalty is a monetary cost calculated
as:
penalty
it
=
P
i
bid
it
P
i
winnerBid
i
P
i
duration
i
(6)
where penalty
it
is the penalty for the i-th CFP due
by provider t, P
i
is the auction’s starting price, bid
it
is
CLOSER2014-4thInternationalConferenceonCloudComputingandServicesScience
44
Figure 2: Auction preparation phase.
the bid called by provider t, winnerBid
i
is the winner’s
bid price and duration
i
is the time frame for which the
computing resource is required by the customer. This
law aims at penalizing the providers proportionally to
their risk attitude. The auction winner who is eventu-
ally unable to meet the request will pay a penalty of
P
i
duration
i
. The following best bidders (2nd, 3rd,
ect..) who on their turn are not able to serve the re-
quest will pay a lower penalty as their bid is higher
than the winner’s. If all participating providers hap-
pen to fail the provision due to the overbooking, the
auction will be closed and a new auction will be called
up.
4 EXPERIMENTAL RESULTS
In order to prove the viability of our proposal a sim-
ulative approach was followed. A prototype was
implemented on the well-known Cloudsim simula-
tor (Calheiros et al., 2011). We developed a new
component (the Auctioneer) and modified the be-
havior of other existing components (Datacenter,
Broker and Cloudlet). Cloudlet is the component
of Cloudsim representing the task submitted by the
customer, while Datacenter is representative of the
provider that will compete for acquiring the task.
Figure 2 shows the basic steps of the auction
preparation phase. The Auctioneer registers to the
Broker (step 1) and receives the list of available Dat-
acenters. When a Cloudlet is submitted to the Broker
(step 2) it is passed on to the Auctioneer (step 3). The
Auctioneer broadcasts a Call For Proposals (CFP)
which specifies the task and the auction requirements
(step 4). Each interested Datacenter checks its own
capacity, reserves a VM and responds to the call (step
5-6). If the Datacenter has run out of physical re-
sources, but is adopting the overbooking mechanism,
it will check out whether the overbooking limit has
been overcome. If not, it will answer the call. If at
Auctioneer
Datacenter1
DatacenterN-1 DatacenterN
Broker
1.StartRound
1.StartRound
1.StartRound
2. CalculateBid
2. CalculateBid
2. CalculateBid
ReturnBid
ReturnBid
ReturnBid
3. SelectWinner
Loop until Max Nr Rounds or No Response
4. NotifyAuctionWin
5. UnableToInstantiate
return FailInstantiationOfVm
6. NotifyPenalty
7. SelectNextWinner
8. NotifyAuctionWin
9. InstantiateVm
Return Vm
10. CommunicateResult
10. CommunicateResult
11. CommunicateBinding
Figure 3: Auction running phase.
least two Datacenters answers the CFP (step 7), the
auction is ready to start.
The auction running phase is depicted in Figure 3.
The Auctioneer notifies the participant that the round
is open (step 1). At the beginning of each round the
starting price is communicated to the participants. In
the very first round the starting price is set equal to
a standard on-demand fixed-price. In the following
rounds the initial price is the price of the winning bid
of the previous round (this applies only to English
auctions). Each Datacenter calculates the bid accord-
ing to its strategy (step 2) and submits it to the Auc-
tioneer, who will appoint the winner (step 3). If the
auction type is multi-round, the bidding procedure is
repeated until the exit condition holds true. For in-
stance, if the auction is of English type it terminates
when no bid has been received within a round. When
the auction is closed, the Auctioneer sends the result
to the winner (step 4) and wait for the Ack. If the
winning Datacenter is not able to serve the request
(step5), it will return a negative Ack. The Auction-
eer will then contact the second best bidder. If again,
the latter has run out of resources, the process will go
further until a positive Ack is returned (step 10) or no
Datacenter is able to honor the contract. In the first
case the auctioneer notifies the losers (step 11) and
the broker that will bind the VM to the winning Dat-
ProcurementAuctionstoMaximizePlayers'UtilityinCloudMarkets
45
Table 1: Weight Setting for the Datacenters’ strategies.
Datacenter ID Strategy w
1
w
2
w
3
w
4
w
5
DC1 Adaptive 0.6 0.1 0.1 0.1 0.1
DC2 Adaptive 0.1 0.6 0.1 0.1 0.1
DC3 Adaptive 0.1 0.1 0.6 0.1 0.1
DC4 Adaptive 0.1 0.1 0.1 0.6 0.1
DC5 Adaptive 0.1 0.1 0.1 0.1 0.6
DC6 Adaptive 0.2 0.2 0.2 0.2 0.2
DC7 Adaptive 0.2 0.2 0.2 0.2 0.2
DC8 Adaptive 0.2 0.2 0.2 0.2 0.2
DC9 Adaptive 0.2 0.2 0.2 0.2 0.2
DC10 Adaptive 0.2 0.2 0.2 0.2 0.2
DC11 Random
acenter (step 12). In the second case the auction is
considered unresolved, and a new auction is called.
Typical Datacenters’ VM allocation policies are:
First fit: a VM is allocated in the first available
host which is capable of running it. According to
this policy, a single physical machine is used to
host VMs until there is available computing ca-
pacity. When that machine capacity is saturated,
a new unused physical machine will be activated.
Worst fit or Balanced allocation: a VM is allo-
cated in the less saturated host among those capa-
ble of running it. This policy ensures a balancing
of the load among the hosts, but causes high frag-
mentation of hosts’ computing capacity.
Best fit: a VM is allocated on the host having the
least amount of unused capacity among those ca-
pable of running it. This policy optimize the re-
source utilization. This is the policy we used for
our tests.
To test the adaptive strategy, we created a set of
11 Datacenters, ten of which adopt the proposed adap-
tive strategy, while one adopts a Random strategy: the
latter makes its bids using the same equation of the
adaptive strategy (Eq. 1), where the α parameter is as-
signed random values in the [0,1] range, without any
specific objective to pursue. The weights characteriz-
ing the α parameter ((Eq. 2) are shown in Table 1. As
the reader may notice, strategies were expressly split
in unbalanced, for which Datacenters point on just
one factor, and balanced, for which all the weights
are assigned the same value.
The objective of the simulation is to show that
strategies actually guide Datacenters in the choice of
tasks to compete for. Every Datacenter counts 70
hosts, each characterized by the following features:
number of cores uniformly chosen between
[64,128,256,512];
RAM: 320 GB;
Storage: 10 TB.
Figure 4: Datacenters Utilization: ER Auction.
A core is modeled in CloudSim with a capacity of
2400 Mips. in the first battery of simulation we sub-
mitted 25000 cloudlets having a length uniformly dis-
tributed in the range [1,24] hours and requiring a VM
type uniformly distributed in the range of the eight
VM types described in Section 3.1. The interarrival
times of the cloudlets are distributed accordingly to
a Poisson distribution, with λ = 0.1 (10 secs is the
maximum time that lapses between the arrivals of two
consecutive cloudlets). From early results, we noticed
that the adaptive strategy of each Datacenter guaran-
tees the achievement of the objective, regardless of
the specific auction type. For this reason (but also for
space reasons) we are going to show the results ob-
tained from testing the English Reverse Auction.
The main parameter we measured was the utiliza-
tion rate of the Datacenters, which is depicted in Fig-
ure 4. We can observe that all the Datacenters (DC)
pursuing one (or a combination) of these two objec-
tives 1) to acquire VMs that require high capacity in
terms of computing resources and 2) to obtain longer
lasting tasks, actually reach an high level of utilization
rate (in Table 1 DC2 and DC4 respectively).
DC11, that adopts the Random strategy, reaches
an high level of utilization rate too, because it can eas-
ily win auctions for tasks that do not meet the objec-
tives of other Datacenters (i.e. low performing VMs,
short tasks).
DC1’s objective is to optimize the utilization rate;
in the graph it can be noticed that after reaching an
utilization rate between 60% and 80%, it is not able
to further increase it, as its objective has almost been
reached: the strategy aggressiveness decreases in a
way that no more auctions are won.
The Datacenter that obtains the lowest utilization
rate (around 20%) is DC3; it exhibits low aggres-
siveness in the auctions as its objective is to obtain
cloudlet with a price not too far from a standard on-
demand fixed-price. However, as it can be seen in Fig-
ure 5 where the revenue loss percentage of the Data-
CLOSER2014-4thInternationalConferenceonCloudComputingandServicesScience
46
Figure 5: Revenue Loss Percentage per Datacenter.
(a)
(b)
Figure 6: Auctions won by DC2 and DC4, grouped by the
length of the Cloudlets the type of VMs respectively.
centers is shown, the objective of DC3 guarantees the
lowest revenue loss. Datacenters with balanced strat-
egy may also avoid revenue losses while, at the same
time, reaching a better utilization rate than DC3.
Finally, we report some graphs showing the
cloudlet characteristics of auctions won by two spe-
cific Datacenters. Figure 6(a) shows the rate of
auctions won by DC2, grouped by the length of
the cloudlets expressed in hours. DC2 mainly won
cloudlets with a length of more than eleven hours (the
reader can check in Table 1 that the weight of param-
eter related to the length of the cloudlets is the high-
est). Figure 6(b) depicts the auctions won by DC4. It
mainly wins auctions requiring high performing VMs,
as it strategy is set to point on that type of VMs.
In order to check the overbooking mechanism a
new simulation battery was run. We created a set of 8
Datacenters enforcing an adaptive strategy with bal-
anced weights, as shown in Table 2. First, the sim-
ulator was fed with 40.000 cloudlets having a length
uniformly distributed in the range [1,12] hours and
poissonian arrivals with λ = 0.01. In Figure 7 we may
notice that all datacenters reach an utilization very
close to 100%. This is due to the fact that cloudlets
are long (in terms of time required by the task) and
Table 2: Weight Setting for the Datacenters’ strategies.
Datacenter ID Strategy w
1
w
2
w
3
w
4
w
5
DC1 Adaptive 0.2 0.2 0.2 0.2 0.2
DC2 Adaptive 0.2 0.2 0.2 0.2 0.2
DC3 Adaptive 0.2 0.2 0.2 0.2 0.2
DC4 Adaptive 0.2 0.2 0.2 0.2 0.2
DC5 Adaptive 0.2 0.2 0.2 0.2 0.2
DC6 Adaptive 0.2 0.2 0.2 0.2 0.2
DC7 Adaptive 0.2 0.2 0.2 0.2 0.2
DC8 Adaptive 0.2 0.2 0.2 0.2 0.2
Figure 7: Datacenters Utilization without overbooking -
40000 Cloudlets, max length = 12h, λ = 0.01.
distant enough (100 secs is the cloudlets’ interarrival
time). We then repeated the simulation with the same
number of cloudlets but with a length uniformly dis-
tributed in [1,6] and poissonian arrivals with λ = 0.06
(17 seconds between consecutive cloudlets). In this
case the average utilization decreases to 70-80%, as
shown in Figure 8.
What happens is that, being the interarrival time
shorter, datacenters simultaneously engage in many
auctions. Each datacenter will only win a subset of
these auctions, which are also very short, thus it will
not be able to saturate its capacity. In this specific case
Figure 8: Datacenters Utilization without overbooking -
40000 Cloudlets, max length = 6h, λ = 0.06.
ProcurementAuctionstoMaximizePlayers'UtilityinCloudMarkets
47
(a)
(b)
(c)
Figure 10: Overbooking Revenues and Penalties.
Figure 9: Datacenters Utilization with overbooking - 40000
Cloudlets, max length = 6h, λ = 0.06.
it may be of help to opt on the overbooking. We then
ran a new simulation, with the same parameters, but
configuring four of the eight datacenters to enforce a
20% of overbooking (DC1,DC2,DC3,DC4). Figure 9
shows that the the overbooking datacenters (depicted
in red) reach a very high utilization (close to 100%).
The collateral effect is of course that they incur penal-
ties, which have been evaluated with the formula in 6.
The revenue of datacenters enforcing the overbooking
drops below the revenue of datacenters which do not
use overbooking, as depicted in 10(a). Put in this way,
there is no point in opting for the overbooking. We
considered new formulas for the evaluation of penal-
ties which consider a reduction of 20% and 30% with
respect to the penalty evaluated in 6. Figures 10(b)
and 10(c) shows that with the new formulas data-
centers enforcing the overbooking have an acceptable
revenue. This is to say that overbooking is an opportu-
nity which must be carefully evaluated by datacenters,
and must be suitably tuned against the penalty policy
adopted by the market. Last consideration is on single
round auctions. Both First and Second Price Sealed
Bid did not provide encouraging results for the over-
booking. This is because the auctions are resolved in
a very short time and the utilization of resources is
not as dynamic as it is in a multi-round auction. We
have then compared the performance of the First and
the Second Price Sealed Bid auctions focusing on the
average utility of the provider. As depicted in Figure
11, the second price auction guarantees, on average,
a better utility. This kind of auction, in fact, let the
datacenter bid its real evaluation of the cloudlet pre-
venting the utility from excessively decreasing.
5 CONCLUSION
Cloud computing has stimulated a great interest both
in the academic community and in business contexts.
More and more IT players look at this technology as
a great opportunity of increasing their profit. Though
several studies report the cloud services’ market rev-
enue is rocketing, economists say the business po-
tential of cloud computing is not yet fully exploited.
There is not yet an open market of cloud resources
where providers and consumers can meet to satisfy
their needs. In this paper we propose a market of
resources where demand and offer of resources can
be matched in auction-based sales. Specifically, we
looked at this market from the perspective of the
provider, who needs a strategy to allocate at best their
unused computing capacity. We proposed an adaptive
strategy that, suitably tailored to the provider’s busi-
ness objective, will help them to maximize the rev-
enue in the context of procurement auctions. Also,
the resource overbooking mechanism has been inves-
tigated as an optional strategy providers may adopt
in order to increase their revenue. Simulations run to
test the proposed approach gave encouraging results,
by showing that each provider is able to reach their
objectives by finely tuning the weights associated to
their strategy. In the future, more factors will be taken
into account in the definition of the provider’s strat-
egy. Further, the business models of the broker of re-
sources (auctioneer) will also be investigated, in order
to prove that a market model based on procurement
CLOSER2014-4thInternationalConferenceonCloudComputingandServicesScience
48
Figure 11: Utility Comparison between FPSB and SPSB.
auctions can yield profit for all market actors.
REFERENCES
Agmon Ben-Yehuda, O., Ben-Yehuda, M., Schuster, A.,
and Tsafrir, D. (2011). Deconstructing amazon ec2
spot instance pricing. In Cloud Computing Technol-
ogy and Science (CloudCom), 2011 IEEE Third Inter-
national Conference on, pages 304–311.
Buyya, R., Ranjan, R., and Calheiros, R. N. (2010). Inter-
cloud: utility-oriented federation of cloud computing
environments for scaling of application services. In
Proceedings of the 10th international conference on
Algorithms and Architectures for Parallel Processing
- Volume Part I, ICA3PP’10, pages 13–31.
Buyya, R., Yeo, C. S., and Venugopal, S. (2008). Market-
oriented cloud computing: Vision, hype, and reality
for delivering it services as computing utilities. In
High Performance Computing and Communications,
2008. HPCC ’08. 10th IEEE International Conference
on, pages 5 –13.
Calheiros, R., Ranjan, R., Beloglazov, A., De Rose, C. A. F.,
and Buyya, R. (2011). Cloudsim: A toolkit for model-
ing and simulation of cloud computing environments
and evaluation of resource provisioning algorithms. In
Software: Practice and Experience.
Chard, K. and Bubendorfer, K. High performance resource
allocation strategies for computational economies. In
IEEE Transactions on Parallel and Distributed Sys-
tems, volume VOL. 24, NO. 1.
Cramton, P., Shoham, Y., and Steinberg, R. (2005). Combi-
natorial auctions. The MIT Press.
Di Modica, G., Petralia, G., and Tomarchio, O. (2013). Pro-
curement auctions to trade computing capacity in the
Cloud. In Eighth International Conference on P2P,
Parallel, Grid, Cloud and Internet Computing (3PG-
CIC 2013), Compiegne (France).
Klemperer, P. (1999). Auction theory: A guide to the liter-
ature. Journal Of Economic Surveys, 13(3).
McAfee, R. P. e McMillan, J. (1987). Auctions and bidding.
Journal of Economic Literature, 15:699–738.
Parsons, S., Rodriguez-Aguilar, J. A., and Klein, M. (2011).
Auctions and bidding: A guide for computer scien-
tists. ACM Computing Surveys, 43(2).
Phillips, R. (2005). Pricing and revenue optimization. Stan-
ford University Press.
Risch, M., Altmann, J., Guo, L., Fleming, A., and Courcou-
betis, C. (2009). The gridecon platform: A business
scenario testbed for commercial cloud services. In
Grid Economics and Business Models, volume 5745,
pages 46–59. Springer Berlin Heidelberg.
Vinu Prasad, G., Rao, S., and Prasad, A. (2012). A combi-
natorial auction mechanism for multiple resource pro-
curement in cloud computing. In Intelligent Systems
Design and Applications (ISDA), 2012 12th Interna-
tional Conference on, pages 337–344.
Wang, Q., Ren, K., and Meng, X. (2012). When cloud
meets ebay: Towards effective pricing for cloud com-
puting. In INFOCOM, 2012 Proceedings IEEE, pages
936–944.
Zaman, S. and Grosu, D. (2013). Combinatorial auction-
based allocation of virtual machine instances in
clouds. J. Parallel Distrib. Comput., 73(4):495–508.
ProcurementAuctionstoMaximizePlayers'UtilityinCloudMarkets
49