Brokering SLAs for End-to-End QoS in Cloud Computing
Tommaso Cucinotta, Diego Lugones, Davide Cherubini and Karsten Oberle
Bell Laboratories, Alcatel-Lucent, Dublin, Ireland
Keywords:
Quality of Service, SLAs, Cloud Computing, Cloud Broker.
Abstract:
In this paper, we present a brokering logic for providing precise end-to-end QoS levels to cloud applications
distributed across a number of different business actors, such as network service providers (NSP) and cloud
providers (CSP). The broker composes a number of available offerings from each provider, in a way that
respects the QoS application constraints while minimizing costs incurred by cloud consumers.
1 INTRODUCTION
Cloud Computing introduces a novel model of com-
puting that brings several technological and business
advantages. Customers can rent cloud services on
a pay-per-use model, without the need for big in-
vestments for resources that have to be designed for
peak workloads, whilst being at risk of remaining
under-utilized for most of the time. Providers may
offer cloud services for rental, hosting them on big
multi/many-core machines, where the infrastructure
investments may be amortized over thousands of cus-
tomers.
However, the requirements of cloud customers
are evolving quickly, as cloud technology is being
massively used worldwide. Many enterprise applica-
tions that might take advantage from the cloud model
cannot be hosted on current infrastructures due to
their stringent performance aspects requiring more
and more from the best effort Internet. Think of vir-
tual desktop, Network Function Virtualization (NFV),
professional on-line multimedia editing and collabo-
rative tools, and on-line gaming, just to mention a few.
Even though recent standards and research efforts
deal with predictable QoS levels in Cloud Comput-
ing (Oberle et al., 2013), spanning across different
business and operations domains remains a great chal-
lenge. For example, a single user request may have
to traverse access, metro, core and data center net-
works, and the top-down provisioning chain across
the various cloud layers (SaaS, IaaS, etc.), and still re-
quire tight interactivity. Figure 1 depicts the situation.
Understanding how to combine all business actors in
agreements to deliver end-to-end QoS may become
overly difficult.
Figure 1: Users crossing multiple network service providers
for accessing a cloud data center.
For these reasons, it is becoming increasingly
important to have available intermediation services
(a.k.a., brokerage) allowing cloud consumers to inter-
face themselves to the multitude of service providers,
including network carriers and cloud providers in-
volved in the end-to-end service delivery chain. Bro-
kering of cloud services (Plummer, 2012) allows for
dealing with the various aspects of SLAs possibly in-
volved in an end-to-end chain, including legal, eco-
nomical and technological aspects. However, in this
work we deal mainly with the latter aspects. That
is, we consider cloud consumers who care about the
price of a service and the expected and received end-
to-end quality.
In this paper, we propose a brokering approach for
610
Cucinotta T., Lugones D., Cherubini D. and Oberle K..
Brokering SLAs for End-to-End QoS in Cloud Computing.
DOI: 10.5220/0004959706100615
In Proceedings of the 4th International Conference on Cloud Computing and Services Science (CLOSER-2014), pages 610-615
ISBN: 978-989-758-019-2
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
deploying distributed cloud applications with guaran-
teed end-to-end QoS which is achieved as an end-to-
end composition of SLAs to be established along the
multiple players participating in the chain. We focus
on the mathematical formalization of the brokering
logic as a mixed-integer geometric programming pro-
gram. Note that, albeit a prerequisite to the brokerage
of cloud services is a certain level of interoperabil-
ity and compatibility among the offerings, we do not
address these specific issues in this paper. Namely,
we assume our brokering logic is managing only ser-
vices and offerings that can interchangeably be used
and composed with each other, for deploying end-to-
end applications. For an overview of the issues behind
cloud interoperability, we refer the reader to (Petcu,
2011).
2 RELATED WORK
The problem of brokerage of multiple CSP offer-
ings for hosting distributed cloud applications with
minimum cost has been addressed in (Houidi et al.,
2011). It has been modeled as a MILP, similarly to
what is done in this paper. However, Houidi’s model
is too simple as it does not consider QoS require-
ments nor capacity constraints. Pawluk et al. de-
fined a cloud broker service (Pawluk et al., 2012)
(STRATOS) based on comparing different CSP of-
ferings. However, the presented broker logic is sim-
ply based on configuration properties matching (e.g.,
minimum values to be satisfied on certain properties
for the offered services). InterCloud (Buyya et al.,
2010) and other architectures (Grivas et al., 2010)
(Ferrer et al., 2012) for federation of cloud providers
include a broker component that may be an excellent
candidate for realizing the brokering logic described
in this paper.
The IRMOS European Project has addressed how
to enhance execution of real-time multimedia applica-
tions in virtualized Cloud infrastructures (Cucinotta
et al., 2010; Kyriazis et al., 2011). IRMOS ensures
resource allocation for individual hosted applications.
However, one of its limitations is that it does not ad-
dress how to possibly guarantee service levels across
domain boundaries (i.e., whenever interacting with
other providers and infrastructures).
The SLA@SOI EU Project
1
developed a frame-
work (Wieder et al., 2011) for negotiation, provi-
sioning, monitoring and adaptation of SLAs through
the entire cloud service life-cycle. It includes both
functional and non-functional characteristics of ser-
1
More information is available at: http://sla-at-soi.eu/.
vices, such as QoS constraints, which can be formal-
ized through an XML-based syntax.
The ETICS (Economics and Technologies for
Inter-Carrier Services) European Project
2
investi-
gated on the critical issues for the creation of a
new ecosystem of innovative QoS-enabled intercon-
nection models between Network Service Providers
(NSPs) impacting all of the actors involved in the end-
to-end service delivery value-chain. ETICS produced
a novel architecture (Zwickl and Weisgrab, 2013) for
control and management of automated end-to-end de-
livery of QoS-enabled services across heterogeneous
carrier networks.
Furthermore, many studies exist trying to identify
critical factors affecting a market of cloud offerings,
including considerations related to QoS-aware provi-
sioning of resources and services. An overview of
such approaches can be found in (Breskovic et al.,
2013).
Concerning standarization, the European
Telecommunications Standards Institute (ETSI)
is currently involved in several activities related to the
issues mentioned above with the ETSI NFV Industry
and Specification group. The Technical Committee
CLOUD is also addressing interoperability aspects
of end-to-end applications. NIST is also working in
this area, with a series of reports being of value to the
topic of end-to-end cloud service quality
3
, in addition
to the well-known reports on the definition of cloud
computing (Mell and Grance, 2011) and its reference
architecture (Liu et al., 2011), where the importance
of having cloud brokerage services is emphasized.
Additional details on research and standardization
efforts in this area can be found in (Oberle et al.,
2013).
3 BROKERING OPTIMIZATION
LOGIC
Notation and Assumptions. From an abstract
viewpoint, a distributed real-time cloud computing
application is modeled as a set of components A =
{
1, 2, . . .
}
, among which we include also client/user-
side components. These are interconnected in a logi-
cal application topology (limited to be a linear work-
flow in this paper) expressed as the sequence S
(s
1
, . . . , s
n
) of n elements of A that activate one after
another, whenever triggered by users requests, com-
ing at a minimum inter-arrival time T. Each activa-
2
More information at: https://www.ict-etics.eu/.
3
More information can be found at:
http://www.nist.gov/itl/cloud/index.cfm.
BrokeringSLAsforEnd-to-EndQoSinCloudComputing
611
Figure 2: Simple application sequence graph.
tion i of component s
i
is associated with its com-
puting requirements C
i
, representing the amount of
time needed for (sequential) computations within the
component for each request received from the user.
After computations, the component sends a message
of size M
i
to the next component s
i+1
in the se-
quence S. For example, consider a cloud application
composed of 3 components A =
{
a, b, c
}
, where a
might represent the client part of the application run-
ning at the consumer premises (e.g., a web browser),
b might represent the front-end server (e.g., a web-
server), and c might represent the back-end compo-
nent (e.g., a database). Then, a sequence might look
like S = (a, b, c, b, a), where user requests traverse
the components forward and backwards.
A cloud application has to be deployed over a set
of cloud data centers P belonging to one or more
Cloud Service Providers (CSPs), interconnected by
means of network providers (NSPs). Each NSP is
characterized by a set of Edge Routers (ERs) R ,
through which it is connected either to other NSPs
or to CSPs or directly to users. From a mathematical
perspective, we consider in what follows data center
locations as edge routers as well, resulting in P R .
The situation can be depicted as a graph where each
node represents an ER which has arcs towards all the
other ERs to which it is connected.
CSPs are assumed to be able to host one or more
components i by reserving a certain amount of com-
puting resources u
i
to it with a minimum VM comput-
ing latency of L
C
, and we assume an ideal scalability
model: if the component is assigned computing ca-
pacity u
i
(0, U
max
] R with VM wake-up latency of
L
C
, then the time needed to complete a user request is
reasonably approximated or bounded as:
C
i
u
i
+L
C
. The
VM computing latency may depend on many factors,
such as hardware and software configuration of hyper-
visors and guest OSes, including the CPU schedulers
type and their configuration.
Similarly, NSPs are assumed to be able to place
the communication flow between each component i
and the next one by dedicating a network bandwidth
b
i
and ensuring a maximum communication latency
L
N
, in such a way that the time needed to transmit
a message among these components (assuming no
queuing of messages from the same application/flow
is needed) can be reasonably estimated/bounded as:
M
i
b
i
+ L
N
.
We assume that our brokering logic has available a
set of quotations (or equivalently a price list) from all
of these providers. More precisely, for each available
data center p P, we have available a few alternate
quotations R
p
. Each quotation r R
p
includes the
cost for deploying components at the CSP premises
that can be made available within conditions of quo-
tation r; this includes K
C
p, r
, the cost of a computing
capacity unit up to a maximum computing capacity
U
C
p, r
, and K
N
p, r
, the cost of a networking capacity unit
up to a maximum networking capacity of U
N
p, r
; the
quotation also includes a maximum VM access la-
tency of L
p, r
. For example, the same CSP might pro-
vide two different quotations with different latencies,
depending on the type of hardware where the VM(s)
would be deployed, and/or the hypervisor and guest
OS kernel configuration(s) , and/or the level of con-
solidation of multiple VMs onto the same host pushed
by the provider in the context of that quotation; also,
the latency might depend on the network configu-
ration within the data center, so the CSP might be
willing to offer VMs at particularly chosen locations
within the DC, and/or using particular configurations
of, e.g., an SDN-enabled data center network, so as to
guarantee lower access latencies from the outside, at
a conveniently higher unit-cost; different quotations
may equally well represent different “sizes” of VMs
that can be seen in current CSPs IaaS offerings. Also,
for each NSP connecting two data centers or edge
routers p, q R , we have available a few alternate
quotations R
p, q
. Each quotation r R
p, q
is expressed
as a cost K
p, q, r
for deploying, with latency L
p, q, r
,
a unit of communication bandwidth between p and
q, up to a maximum available bandwidth of B
p, q, r
.
In cases where, among two data centers p, q P ,
there are multiple NSPs connecting them, or there are
multiple paths traversing different NSPs for connect-
ing them, we assume that all the corresponding quo-
tations are combined together into a single list R
p, q
combining together the available multiple quotations
to traverse. For example, quotations from two NSPs
to be traversed for connecting two locations p, q P
can be combined as follows: for each pair of quo-
tations r
1
from the first NSP and r
2
from the sec-
ond NSP, we consider an equivalent quotation r
r
1
, r
2
in which the bandwidth-unit costs and the latencies of
r
1
and r
2
area added together, whilst the maximum
available bandwidth for the quotation is the minimum
among the two. After this operation of aggregating
quotations from different NSPs, we obtain equivalent
end-to-end quotations for all pairs of data center lo-
cations p, q P , thus in our formalization we can
forget about the existence of the intermediate edge
routers R \P among multiple NSPs carrying the traf-
fic among all CSPs.
CLOSER2014-4thInternationalConferenceonCloudComputingandServicesScience
612
Problem Formalization. Now we proceed to for-
malize the cloud brokering problem as an optimiza-
tion program. To this purpose, we introduce Boolean
variables x
i, p, r
representing whether the broker will
deploy component i onto data center location p P
making use of quotation r R
p
, and the Booleans
y
i, p, q, r
representing whether or not traffic between
components s
i
and s
i+1
in sequence S is to be de-
ployed between locations p, q P by using quotation
r R
p, q
. These variables are tied to each other by the
constraints:
pP
rR
p
x
i, p, r
= 1 s
i
S
(1)
rR
p, q
y
i, p, q, r
=
r
1
R
p
x
s
i
, p, r
1
!
r
2
R
q
x
s
i+1
, q, r
2
!
p, q P , s
i
S\
{
s
n
}
(2)
where: the first constraint forces each component
to be placed onto a single cloud provider, exploiting
one among the available quotations; the second one
says that, if component s
i
is deployed on p, and com-
ponent s
i+1
on q, then only one among the available
quotations for communications between p and q has
to be chosen; otherwise, y
i, p, q, r
must be 0 for each
r R
p, q
. Also, the deployment and quotations choice
has to satisfy the overall end-to-end deadline D con-
straint:
s
i
S
C
i
u
i
+
pP
rR
p
L
C
p, r
x
i, p, r
+
M
i
b
i
+
p, qP
rR
p, q
L
N
p, q, r
y
i, p, q, r
!
D (3)
and the allocation constraints
s
i
S
u
i
x
i, p, r
U
p, r
p P , r R
p
(4)
s
i
S
b
i
y
i, p,q,r
B
p,q,r
p, q P r R
p, q
(5)
where u
i
and b
i
are real problem variables. These
are further constrained due to the assumption that re-
quests submitted to the same application workflow at
the specified minimum inter-arrival time of T do not
queue after each other, namely:
u
i
C
i
T
(6)
b
i
M
i
T
. (7)
On the other hand, the broker will try to minimize
the cost for hosting the application:
min
s
i
S
u
i
pP
rR
p
K
p, r
x
i, p, r
+ (8)
s
i
S\
{
s
n
}
b
i
p, qP
rR
p, q
K
p, q, r
y
i, p, q, r
. (9)
The formalized problem falls in the class of mixed-
integer geometric programming optimization pro-
grams, for which there are solvers available, such
those found in the GAMS suite
4
.
4 IMPLEMENTATION
CONSIDERATIONS
Network Service Provider. In order to provide
end-to-end cloud services with precise QoS levels,
NSPs and CSPs must implement selected routing
paths and reserve appropriate resources. On the net-
working side, we distinguish between intra-domain
routing if the end-to-end path is entirely within the
same Autonomous System (AS), and inter-domain
routing if the traffic is sent across multiple NSPs.
In intra-domain routing, users and DCs where
their applications are deployed are all interconnected
via the same NSP. In order to compose different net-
working offers, the NSP can adopt different strategies
to build the end-to-end path. For example, a best ef-
fort latency offer can be realized by adopting tradi-
tional IGP protocols (e.g., OSPF or IS-IS) to route
traffic between source and destination on the shortest
path. Of course, not always shortest path translates
into minimum latency. On the other hand, a premium
offer (or minimum latency offer) can be realized us-
ing Multi Protocol Label Switching (MPLS) with its
Traffic Engineering extensions (MPLS-TE). In partic-
ular, a Label Switched Path (LSP) is a one-way tun-
nel that can transport the traffic from origin to desti-
nation and that can satisfy multiple QoS parameters
(e.g., bandwidth, delay, jitter, availability, and loss).
The LSP routing path can either be configured auto-
matically (e.g., using Constrained Shortest Path First
CSPF), or manually, e.g., as a result of more sophis-
ticated techniques (Cherubini et al., 2011).
In inter-domain routing, when users can reach
CSP’s DCs across different NSPs, optimal path calcu-
lation is inherently more complex. Each router should
have a global view of the network. Unfortunately,
NSPs “filter” important information needed for the
path establishment (e.g., for scalability or confiden-
tiality reasons). The Path Computation Element pre-
4
More information is available at:
http://www.gams.com/solvers/
BrokeringSLAsforEnd-to-EndQoSinCloudComputing
613
sented in (Farrel et al., 2006), represents a possible so-
lution to provide an optimal inter-domain routing path
(in the form of MPLS-TE LSP) that meets desired
QoS requirements and can scale the end-to-end net-
work up to 100,000 MPLS devices (Leymann et al.,
2013).
Cloud Service Provider. The centralized design of
today’s data centers offers advantages in terms of fab-
ric homogeneity and control, in comparison to the
best-effort networks describedabove, which usually
include various legacy equipment, specialized boxes,
multiple protocols and are potentially operated by in-
dependent NSPs.
However, meeting the latency guarantees required
by demanding Cloud applications in the data cen-
ter fabric is still a significant challenge. The shared
nature of the network in multi-tenant data centers
leads to significant variations in the perceived perfor-
mance. The lack of predictability increases the ten-
ant costs and causes provider revenue loss. Moreover,
the Internet-oriented protocols and flow scheduling
mechanisms used in data centers are unaware of flow
deadlines and application traffic patterns. Instead,
they strive to optimize such low-level metrics as net-
work throughput and fairness, ignoring the actual per-
formance requirements associated to traffic flows.
Another reason that makes it difficult to deliver
precise QoS levels in data centers is that typical de-
signs focus on bisection bandwidth by overprovision-
ing the fabric. However, these are clearly not op-
timized for ultra-low latency applications with pre-
dictable delivery of packets. Moreover, overprovi-
sioning today’s data centers is prohibitively expen-
sive.
The most common strategy is to implement
QoS mechanisms by segregating traffic into differ-
ent classes to provide isolation and enable traffic en-
gineering. Typically, these mechanisms are imple-
mented in switches and network cards where traffic
is prioritized explicitly by marking packets or implic-
itly by using port ranges. In addition, the cloud ser-
vices are typically run at low utilization to meet strict
SLA demands, which decreases efficiency. However,
advanced research in the field proposes different alter-
natives such as removing the kernel and network stack
from the critical path of communication and load bal-
ancing requests across application instances (Kapoor
et al., 2012). In (Ballani et al., 2011) authors pro-
pose to create virtual network abstractions to allow
tenants to expose their network requirements. Pro-
posal in (Alizadeh et al., 2012) run the network with
near zero queueing and adaptively respond to conges-
tion marks using DCTCP (Alizadeh et al., 2010).
Also, deadline-based scheduling of packets has
been proposed as an alternative to TCP. In (Wilson
et al., 2011), a control protocol is proposed that uses
deadlines to achieve informed allocation of network
bandwidth. In (Andrews, 2000), EDF scheduling is
leveraged for providing probabilistic end-to-end guar-
antees to individual streams.
5 CONCLUSIONS AND FUTURE
WORK
In this paper, we addressed optimum brokering of
cloud and carrier services accounting for end-to-
end QoS requirements of cloud applications, and the
availability of multiple offerings by CSPs and NSPs
with different and heterogeneous quality levels corre-
sponding to different price conditions. The problem
has been mathematically modeled as a mixed-integer
geometric programming optimization program that
can be solved by available standard solvers. This
work constitutes a first step towards the realization
of evolved end-to-end cloud services offering to con-
sumers precise, stable and reliable QoS levels across
the heterogeneous, multi-provider “supply” chain that
is involved in the process.
Our planned and ongoing future work in the area
includes development of simplified versions of the
introduced problem, as well as fast heuristic solvers
for trading off accuracy of the solution versus solv-
ing time. Also, an implementation of the described
technique is under way and its effectiveness will soon
be evaluated by simulation, within our CloudNetSim
framework (Cucinotta and Santogidis, 2013), and by
real prototyping.
REFERENCES
Alizadeh, M., Greenberg, A., Maltz, D. A., Padhye, J., Pa-
tel, P., Prabhakar, B., Sengupta, S., and Sridharan, M.
(2010). Data center tcp (dctcp). SIGCOMM Comput.
Commun. Rev., 41(4):–.
Alizadeh, M., Kabbani, A., Edsall, T., Prabhakar, B., Vah-
dat, A., and Yasuda, M. (2012). Less is more: trad-
ing a little bandwidth for ultra-low latency in the data
center. In Proceedings of the 9th USENIX conference
on Networked Systems Design and Implementation,
NSDI’12, pages 19–19, Berkeley, CA, USA. USENIX
Association.
Andrews, M. (2000). Probabilistic end-to-end delay bounds
for earliest deadline first scheduling. In INFOCOM,
pages 603–612.
Ballani, H., Costa, P., Karagiannis, T., and Rowstron, A.
(2011). Towards predictable datacenter networks.
SIGCOMM Comput. Commun. Rev., 41(4):242–253.
CLOSER2014-4thInternationalConferenceonCloudComputingandServicesScience
614
Breskovic, I., Brandic, I., and Altmann, J. (2013). Max-
imizing liquidity in cloud markets through standard-
ization of computational resources. In Proceedings
of the 2013 IEEE Seventh International Symposium
on Service-Oriented System Engineering, SOSE ’13,
pages 72–83, Washington, DC, USA. IEEE Computer
Society.
Buyya, R., Ranjan, R., and Calheiros, R. (2010). Inter-
cloud: Utility-oriented federation of cloud computing
environments for scaling of application services. In
Hsu, C.-H., Yang, L., Park, J., and Yeo, S.-S., editors,
Algorithms and Architectures for Parallel Processing,
volume 6081 of Lecture Notes in Computer Science,
pages 13–31. Springer Berlin Heidelberg.
Cherubini, D., Fanni, A., Mereu, A., Frangioni, A., Mur-
gia, C., Scutell
`
a, M. G., and Zuddas, P. (2011). Linear
Programming Models for Traffic Engineering in 100%
Survivable Networks under Combined IS–IS/OSPF
and MPLS-TE. Computers & Operations Research,
38(12):1805–1815.
Cucinotta, T., Checconi, F., Kousiouris, G., Kyriazis, D.,
Varvarigou, T., Mazzetti, A., Zlatev, Z., Papay, J.,
Boniface, M., Berger, S., Lamp, D., Voith, T., and
Stein, M. (2010). Virtualised e-learning with real-
time guarantees on the irmos platform. In Proceed-
ings of the IEEE International Conference on Service-
Oriented Computing and Applications (SOCA 2010),
pages 1–8, Perth, Australia.
Cucinotta, T. and Santogidis, A. (2013). Cloudnetsim -
simulation of real-time cloud computing applications.
In Proceedings of the 4th International Workshop on
Analysis Tools and Methodologies for Embedded and
Real-time Systems, Paris, France.
Farrel, A., Vasseur, J.-P., and Ash, J. (2006). RFC 4655
A Path Computation Element (PCE)-Based Architec-
ture.
Ferrer, A. J. et al. (2012). Optimis: A holistic approach to
cloud service provisioning. Future Generation Com-
puter Systems, 28(1):66 – 77.
Grivas, S., Kumar, T., and Wache, H. (2010). Cloud bro-
ker: Bringing intelligence into the cloud. In Cloud
Computing (CLOUD), 2010 IEEE 3rd International
Conference on, pages 544–545.
Houidi, I., Mechtri, M., Louati, W., and Zeghlache, D.
(2011). Cloud service delivery across multiple cloud
platforms. In Services Computing (SCC), 2011 IEEE
International Conference on, pages 741–742.
Kapoor, R., Porter, G., Tewari, M., Voelker, G. M., and
Vahdat, A. (2012). Chronos: predictable low latency
for data center applications. In Proceedings of the
Third ACM Symposium on Cloud Computing, SoCC
’12, pages 9:1–9:14, New York, NY, USA. ACM.
Kyriazis, D., Menychtas, A., Kousiouris, G., Oberle, K.,
Voith, T., Boniface, M., Oliveros, E., Cucinotta, T.,
and Berger, S. (2011). A real-time service oriented
infrastructure. GSTF International Journal on Com-
puting, 1(2).
Leymann, N. et al. (2013). draft-ietf-mpls-seamless-mpls-
04 – Seamless MPLS Architecture.
Liu, F., Tong, J., Mao, J., Bohn, R., Messina, J., Badger, L.,
and Leaf, D. (2011). Nist special publication 500-292
– nist cloud computing reference architecture.
Mell, P. and Grance, T. (2011). NIST SP800-145: The NIST
Definition of Cloud Computing.
Oberle, K., Cherubini, D., and Cucinotta, T. (2013). End-
to-end service quality for cloud applications. In
Proceedings of the 10th International Conference on
Economics of Grids, Clouds, Systems and Services,
Zaragoza, Spain.
Pawluk, P., Simmons, B., Smit, M., Litoiu, M., and
Mankovski, S. (2012). Introducing stratos: A cloud
broker service. In Cloud Computing (CLOUD), 2012
IEEE 5th International Conference on, pages 891–
898.
Petcu, D. (2011). Towards a Service-Based Internet,
volume 6994 of Lecture Notes in Computer Sci-
ence, chapter Portability and Interoperability between
Clouds: Challenges and Case Study, pages 62–74.
Springer Berlin Heidelberg.
Plummer, D. (2012). Gartner, inc. – cloud services broker-
age: A must-have for most organizations.
Wieder, P., Butler, J., Theilmann, W., and Yahyapour, R.
(2011). Service Level Agreements for Cloud Comput-
ing, springer edition.
Wilson, C., Ballani, H., Karagiannis, T., and Rowtron, A.
(2011). Better never than late: meeting deadlines in
datacenter networks. SIGCOMM Comput. Commun.
Rev., 41(4):50–61.
Zwickl, P. and Weisgrab, H. (2013). Final ETICS architec-
ture and functional entities high level design.
BrokeringSLAsforEnd-to-EndQoSinCloudComputing
615