SLA data management criteria

SLA data management criteria
Katerina Stamou, Verena Kantere, Jean-Henry Morin
Institute of Services Science, University of Geneva, Switzerland
Abstract—Service Level Agreements (SLAs) represent service
management contracts that are processed by monitoring and
measurement mechanisms for the evaluation of the signatories
adherence to the agreed service levels during service execution.
The paper discusses SLA data management characteristics that
need to be considered in the design of data models for SLA
documents. The SLA anatomy is introduced with respect to the
Web Service Level Agreement (WSLA) [1] language specification.
Furthermore, the paper highlights current obstacles for the
integration of automated SLA management in the cloud business
setting. The contributed SLA data analysis maps SLA terms
to data management attributes according to their operational
relevance during the SLA activity. We present an SLA digraph
model for the automated SLA formulation and data handling.
The SLA digraph is introduced as a programming module that
sits on the application layer and communicates with backend
data stores for the SLA persistence.
I. INTRODUCTION
In the cloud computing context, IT services are provisioned
on-demand by utilizing distributed computing resources and
economies of scale. Every customer needs to agree with a
Service Level Agreement (SLA) in order to lease a new
service. Traditionally, providers define SLAs, in which they
guarantee explicit service-level bounds over a predefined,
agreed period. SLAs represent contractual terms and condi-
tions between service providers and customers. Their content
assures the mutually agreed service levels between a provider
and a consumer [2].
SLA contracts describe obligatory service provisioning
terms. Such terms encapsulate Quality of Service (QoS) at-
tributes and functional service properties. Furthermore, SLAs
may include a multitude of technical and business service-
level objectives (SLOs) along with metrics and service pa-
rameter endpoints. Thus, SLAs assist with the calculation and
measurement of service parameters, which in turn indicate the
provider’s adherence to the promised service levels.
We assume that SLAs come in machine-readable format
and that cloud service providers use automated systems for
the management of their offered services. Our view on SLAs
exceeds their basic usability as service level instruments.
We consider SLA templates as what-you-see-is-what-you-get
(WYSIWYG) artifacts that customers can use to negotiate
and finalize their service selection. SLA templates represent
pre-agreed SLAs that typically describe all available service
provisioning options. Moreover, we assume that electronic
marketplaces and service agencies utilize SLAs and SLA
templates to offer customers service-management facilities,
where shared SLA data are distributed over remote repositories
and processed automatically.
The scope of this work is to introduce SLA data char-
acteristics that impose special storage and data operation
structures for SLA documents. Compared to other types of
electronic contracts and software licenses, the values of SLA
terms usually represent the outcome of multiple backend
resource management operations. SLA information needs to
be accessed simultaneously by diverse processes (e.g. moni-
toring, auditing mechanisms) that manage the overall service
execution. An appropriate SLA data structure and persistence
mode can contribute to efficient service level management.
Currently SLAs for cloud services lack standardization that
would provide them with more structured content. Thus, they
mostly appear as semi-structured, static information and their
utilization by virtual cloud markets is limited. Moreover,
the description of offered cloud services is superficial and
the provided SLA contracts are inadequate for customers to
manage their leased services. Given such facts, it is ambiguous
if cloud service customers actually get the service levels that
they are paying for. Our work highlights current obstacles
that restrict the SLA role in service marketplaces to static,
terms-and-conditions documents and augments on research
challenges that can reverse the SLA utilization into a more
dynamic and value-added one.
The paper initially provides a precise analysis of the SLA
structure and primary components according to the WSLA
language [1] and to the WS-Agreement initiation protocol [3].
To avoid a semi-structured, insufficient SLA representation,
the SLA anatomy, formalization and initialization cycle are
discussed with respect to the role of language contractual
elements including signatory parties and by following the
WSLA specification.
The primary contribution of this work is a structured
analysis of SLA data management needs. We organize our
study by collecting a set of generic data operational attributes.
SLA terms are mapped to such attributes according to their
operational relevance in the SLA execution. The analysis helps
to formulate a view on the data structure that is adequate
for SLA data management over distributed web resources.
Moreover, the provided study can be used as a generic tool
for the design and deployment of SLA data stores.
Given the SLA anatomy and data analysis, we use the
WSLA language to build and contribute an SLA digraph data
model, where nodes and edges represent the SLA content.
We propose that the SLA digraph can be used as a data
model for the management of SLA information. The initial
model, presented in this work, is implemented as a Python
2.7 programming module.
The paper is organized as follows: Section II examines the

SLA structure and initialization cycle. Section III discusses
current SLA obstacles that bound the SLA utility in cloud
marketplaces as well as research challenges to overcome such
obstacles. Section IV focuses on the SLA data analysis with re-
spect to data management characteristics. Section V introduces
the proposed digraph model for the data management of SLAs.
Finally, Section VI elaborates on criteria and objectives of our
experimental study with respect to the SLA digraph model
evaluation. We summarize the paper with related research,
conclusions and on-going work.
II. SLA ANATOMY AND FORMALIZATION
The distributed computing community has driven research
and technical advancements on SLA management to cover
immense needs for resource reservation and monitoring of job
execution. The WSLA [1] and WS-Agreement [3] language
specifications have been used by researchers to express SLAs
in a machine-readable format.
The GRAAP working group proposed the WS-Agreement
as a language and a protocol to conduct SLAs. The WSLA
language has been led by IBM research on utility computing.
According to [1], an SLA complements a service description.
An important attribute of both language specifications is that
they enable the automated parsing and manipulation of the
SLA document by backend processing systems.
In contrast to other contracts for IT services, properties
within an SLA have to be monitored during workload ex-
ecution to audit potential service-level violations and verify
adherence to the agreed SLOs. In the literature, SLAs are
often perceived as automated processes that merely assist the
monitoring and scheduling of computational tasks.
Both WSLA and WS-Agreement specifications use a tree
data structure to represent the SLA information. Tree branches
illustrate separate SLA sections and tree leaves inner section
terms. XML is used by both schemas. Figure 1 illustrates a
high level view of the WSLA language anatomy, according to
which the core elements of an SLA are the following:
Fig. 1: Web Service Level Agreement (WSLA) structure
Parties: Signatory parties consist of one service provider and
one service customer. Supporting parties represent third parties
that operate on behalf of either or both signatories.
Service definition: Service objects represent description terms
and include SLA parameters that indicate quantitative as well
as qualitative metrics.
Obligations: A provider defines guarantees in the form of obli-
gations either as service level objectives (SLOs) or as action
guarantees. SLOs represent measurable targets that a provider
promises to fulfill during service execution. SLO values can be
verified via monitoring and auditing. Action guarantees signify
tasks that the provider, one or more supporting parties or, in
some cases, the customer will take to establish the promised
service levels of one or more SLA parameters. A well-defined
description of service and resource parameters is a prerequisite
for the specification of QoS attributes in the form of either
SLOs or action guarantees.
The core elements of the language are further divided into
granular sub-elements. In the case of service objects, the
granularity of information reaches the level of URI sources
that can be monitored and measured. In the case of service
obligations, the flow of the information content involves busi-
ness level objectives, as well as SLO evaluation functions and
expressions that typically follow first order logic.
Moreover, obligations may include the formalization of
penalties in case of service-level violations, as well as of re-
imbursement conditions and rewards that represent penalty
complements. The granularity of guarantee description de-
pends on the information content and on the complexity of the
guarantee evaluation. A guarantee can be defined for a service
atom (i.e. a standalone service or service component) or for a
service bundle. Moreover, a guarantee may be generic enough
and apply to the overall service provisioning (e.g. customer
authentication/authorization options, etc).
A. SLA initiation cycle
In addition to the SLA language structure and semantics
that both specifications [1], [3] provide, the WS-Agreement
defines a protocol for the SLA initiation between contracting
parties. Figure 2 summarizes the basic steps in the information
exchange:
Fig. 2: SLA initialization lifecycle
1) A provider depicts their provisioning availability, capacity
and service leasing options in one or more SLA tem-
plates. An SLA template represents a pre-instantiated
agreement that is prepared by a service provider for
customer consideration. The SLA template describes the
agreement content that a provider is willing to accept
during communications with customers. Thus, it provides
a holistic view on a provider’s resource capacity and
provisioning plan.
2) Both specifications suggest the use of customer and
provider templates for the exchange of counter offers in

the process of agreeing on service levels. SLA templates
are typically accessible to customers through a service
registry that pulls data from designated repositories. To
decide which provisioning is more suitable for their
needs, customers review SLA templates and proceed with
either agreement initiation or negotiation with one or
more providers.
3) After reviewing SLA templates, a customer will select
and customize the one that best suits their computational
needs and will send it back to the service provider as an
agreement request.
4) On receiving the agreement request, a provider will
review it and estimate whether the request falls within
the existing provisioning capacity.
5) After reviewing the agreement request, a provider can
either accept it and initiate the agreement or negotiate
with the customer on service-level terms or reject the
request.
6) Upon agreement initiation, the SLA is subject to the
agreed monitoring and auditing services until the exe-
cution of the computational task is complete.
III. SLA ROLE IN VIRTUAL ECONOMIES
As described in Section II, SLAs control the measurement
and auditing of available resource parameter values. The SLA
definition provides an explicit view on how the provisioning of
a service is planned. It also indicates precise bounds of service
levels that a provider can supply. Providers use SLAs to control
the levels of resource consumption and service availability.
Currently, SLAs hardly appear in cloud marketplaces. The
promotion of IT offers to customers primarily relies on high-
level service descriptions that are typically static and that do
not allow for any processing. The role of SLAs is peripheral
and they are often materialized by documents of ”terms-
and-conditions” that usually do not involve functional service
aspects. On top, SLAs lack standardization, which would in-
evitably lead to more structured, rule-based SLA formulation.
We view SLAs and their derivatives as dynamic informa-
tion that is updated at frequent time intervals. In [4] we
demonstrated how SLA templates can be utilized as customer
drivers for service selection by a third-party service (e.g.
virtual marketplace, service-aggregator) since SLA templates
enclose all details on how services are to be provisioned.
Our experimentation showed that the SLA content modularity
enables the rapid retrieval of all inclusive information through
queries that are performed in one or more databases. This fact
allows for the SLA manipulation in multiple ways, which in
[4] we exercised by treating SLA template sections as facets.
A. Cloud SLAs challenges
In the scientific literature, SLAs are hardly viewed as
end-user documents, but merely as automated processes. In
contrast, cloud marketplaces treat SLAs as static documents.
It is thus challenging to find the right equilibrium between
these two orthogonal aspects and combine machine-readable
with user-friendly SLAs into a uniform process that can be
used by both backend systems and front-end web services.
In the cloud business setting, diversified services are offered.
Description characteristics and provisioning guarantees vary
considerably, even if they describe similar services in different
contexts. Providers from various business domains use cus-
tomized terminology to describe service parameters, metric
functions and guarantee definitions. In service descriptions,
terms like ”availability”, ”throughput” and ”performance” are
usually included in ambiguous ways, which may be confusing
for service customers. Various vocabularies of provisioning
terms represent a primary cause for SLA heterogeneity.
On a wide scale, SLA semantic and structural heterogene-
ity represents a challenge because it complicates the SLA
template comparison and hinders any attempt to efficiently
manipulate SLAs in distributed service markets. SLA con-
tent heterogeneity addresses issues that deal with research
opportunities for data management, information retrieval and
language processing. Moreover, it highlights the need for
SLA standardization that would allow for classification of key
performance indicators (KPIs) or would mandate the inclusion
of specific functions per business domain.
In the cloud computing setting, SLA formulation highly
depends on resource availability that is prone to customer-
demand variations [5]. Hence to manipulate SLA templates,
the template content must be processed dynamically. As the in-
formation depth of SLAs is unbounded, frequent data updates
may cause performance delays in the data exchange between
customers and providers. Yet, a reasonable storage schema
for SLA templates is under question. On one hand, one may
argue that since SLA templates represent dynamic information
objects, they should not be stored at all. Instead, they should
be kept in-memory for as long as they are valid and then
be immediately replaced. On the other hand, a modular SLA
data model can help with the preservation of SLA templates
for longer time periods. Moreover, it can facilitate frequent
content updates according to provider resource capacity and
service availability. Our research centers on the latter aspect.
According to [3] and [1], SLAs represent nested tree struc-
tures. On the business setting, SLA documents may include
heterogeneous characteristics and they are typically unbounded
in terms of length and content. A challenge for the SLA
manipulation is to select a data structure that facilitates quick
traversing within nested information paths. For example, a
modular structure provides isolation between SLA compo-
nents and allows for categorization of SLA terms. Moreover,
an efficient data structure enables the exploitation of finer
grained information, while preserving the necessary SLA
data interconnectivity. Finally, an adequate SLA data model
applies to diverse types of SLAs and can be accomodated
by various management schemas according to application and
infrastructure specifics. The rest of the paper concentrates on
SLA data requirements that indicate a satisfactory SLA data
structure for the cloud computing setting.

IV. SLA DATA ANALYSIS
Under the term SLA data management we enclose all
SLA data operations that take place before, during and after
service execution. Such data include pre-instantiated, active
and terminated SLAs as well as fine-grained service elements
(e.g. resource metrics, service descriptions and provisioning
guarantees), whose combination assembles the SLA content.
In a service economy, the SLA formalization using a
structured, well-defined schema (like [1], [3]) extends SLA
manipulation opportunities towards business-oriented opera-
tions. However, an SLA data model is required to support
the automated management of processes with respect to real-
time data monitoring and auditing. Compared to other types of
service contracts (e.g. terms-and-conditions, software licences)
the values of SLA terms need to be monitored and measured
during service execution to verify that SLOs are met and that
no service violations have occured.
Accordingly, the measurement of data values has to be
depicted in the SLA content in frequent time intervals. The
requirement for real-time data updates particularly applies in
the cloud computing setting, where services are exchanged
on demand and business relationships may enclose financial
responsibilities (e.g. violations of SLA terms may trigger
customer re-imbursements or impair a service provider’s rep-
utation). Furthermore, nested SLA information may include
interdependencies between diverse components or component
sets (e.g. a change in an SLA parameter value may affect re-
spective SLO values). Such dependencies need to be addressed
by the SLA data model as they relate to data correctness.
To support our claim that a graph data model is adequate for
the SLA data management needs, we collected generic, data-
operational attributes and mapped their significance to core
SLA components that are typically met in any SLA. Table I
illustrates this mapping. The first column of Table I lists the
selected data management characteristics. SLA components
are mapped to data attributes according to their required
operations (e.g. monitoring, auditing, alerting) for successful
service management. The last column of Table I illustrates the
significance of data attributes to the overall SLA document.
Since SLAs are not yet standardized and their content is semi-
structured, a multitude of SLA terms may have similar or
additional management needs, which can be determined by
specific application criteria and desired functionality.
Data volume: refers to the size of information to be stored
and processed. In the case of SLAs and SLA components,
the size is negligible, thus there is no direct need for storage
allocation or big data transfers. The volume per service object
(assuming the hierarchical inclusion of SLA parameters, com-
posite and resource metrics) as well as the size of data per
guarantee definition typically lies in the kilobyte range. More-
over, SLA data operations (e.g. customer or provider alerts)
can occupy temporal space instead of disk storage. On the
other hand, and assuming the exploitation of SLAs by public
marketplaces, there can be massive volume of interactive client
requests over stored SLA data. Thus, a flexible data structure
Data charac-
teristics
ST1 ST2 ST3 ST4 SLA
accessibility,
integrity
? ? ? ? ?
velocity rate high high high low na
replication,
staging
? ? ?
service depen-
dencies
? ? ? ?
cleanness ? ? ? ? ?
accuracy ? ? ? ? ?
ownership, au-
thenticity
? ? ? ? ?
heterogeneity ? ? ? ? ?
TABLE I: Data management characteristics- SLA terms map-
ping. ST1: SLA parameter, ST2: composite or resource metric,
ST3: service level objective (SLO), ST4: action guarantee,
SLA: complete SLA document, na: non applicable.
and database schema should allow for alternative persistence
modes that efficiently address dynamic SLA data needs and
that allow for parallel processing of real-time data queries. As
there is no direct mapping with the depicted SLA information,
this attribute is not illustrated in Table I.
Data accessibility, integrity: The definition of resource and
composite metrics typically involves numerous data sources.
Composite metrics can be calculated by the accumulation of
resource metrics from distributed access points (e.g. URIs,
query interfaces). SLA parameters are computed using com-
posite metric values. Typically, a logical expression asserts the
value range that specifies an SLA parameter. Data accessibility
and data-value integrity are of great importance for the correct
processing of SLA templates . The information description has
to be granular enough to enclose metric dependencies, service
access points (SAP) and unique resource identifiers (URIs).
Such information enables resource and service management
operations, e.g. monitoring processes that control the resource
consumption and inform scheduling processes on scale in/out
needs, auditing processes that trigger alerts on possible service
level violations, etc.
Data velocity rate: SLAs contain terms whose values may
change dynamically during service execution. Such terms are
typically represented as resource, composite metrics and SLA
parameters, which are assembled under the definitions of
service objects and represent the holistic service description.
Moreover, the information included in the SLA service/re-
source description is affected by resource demand. As ob-
served by [5] there are time intervals when the consumption
load increases and thus a provider has to either scale out or,
vice versa, apply a more stringent provisioning plan. SLA
guarantees depend very much on the value ranges of SLA
parameters, thus are also volatile and subject to changes
according to resource consumption and availability. Thus, to
support value-added business operations the SLA data schema

must allow for rapid information retrieval and volatile data
management.
Data replication, staging: SLA parameter and metric val-
ues may affect the definition of other SLA terms, like SLOs
or action guarantees. In the cloud setting, such information is
accumulated or processed directly from distributed resources,
thus data may have to move among diverse repositories or to be
transformed into different formats. With respect to SLA guar-
antees, data staging and replication may depend from provider-
customer arrangements and from the provisioned service type
or SLA class. For example, an agreed SLA may consist
of smaller SLAs that have limited time-spans and combine
schedules of parallel and sequential service execution tasks.
On completion of an SLA subset, active SLA information may
have to be replicated or to migrate into new data repositories
where a different data format might be expected.
Service inner-dependencies: The evaluation of SLA pa-
rameter values may depend on the measurement of one or
more metrics. With the term service dependency the authors
in [6] define the relationship between an antecedent service or
application component that requires an operation performed by
a precedent component in order for the former to execute its
function. The authors use the terms antecedent and precedent
to express the two counterparts of the dependent relationship.
By default such relationship is directed.
The authors in [6] differentiate service dependencies be-
tween functional and structural. The former represents generic
service dependencies and principal service components to
which all other service models are bound (e.g. service type,
service name, service purpose (end-user, backend)). Structural
dependencies illustrate detailed descriptions of components,
whose interoperation represents the offered service. The struc-
tural model complements the functional one and provides ex-
plicit descriptions of service elements (technical, configuration
installations etc). SLA component inner-dependencies need to
be reflected in the SLA data structure and database schema
since they affect operations that influence the overall SLA
processing.
Data cleanness: With respect to cloud services, the SLA
service/resource description may define the maximum number
of database instances, tenants or disk space that is allocated per
virtual machine (VM) and VM instance; moreover it may de-
scribe the virtualization platform and CPU characteristics like
processor power and type. The SLA service description depicts
the overall provisioning capacity, thus resource availability.
Such data along with SLA guarantees should be publicly
readable to help customers project their service requirements
given available resources. Thus, elimination of data noise and
preservation of data cleanness represent necessities for the
processing of SLA information.
Data accuracy: As guarantee information enables both
provider and customer to evaluate the levels of service pro-
visioning, the provider needs to ensure the information ac-
curacy in terms of real-time updates and correct resource
measurement. Data conformity helps the service provider to
preserve promised quality-of-service (QoS) levels and assures
customers with respect to service consumption.
Data authenticity, ownership: During the SLA initiation
cycle (Fig. 2), SLA information is exchanged and modified
by the prospective signatories. Thus, a mechanism needs to
guarantee the data transfer integrity and privacy among the
contracting entities. Moreover, providers formulate guarantees
that define value range limits for SLA parameters and con-
figuration options for service objects. As the reputation and
business prosperity of a provider primarily depends on the
accomplishment of promised and agreed service levels, the
provider has to ensure the authenticity, non-repudiation and
ownership of publicly readable SLA information.
Data heterogeneity: As discussed in Section III, cloud
SLAs lack standardization that permits a more structured
content and data schema. Additionally, the variety of services
offered in the cloud context requires alternative SLA forms
and compliant formats that can be processed by existing web
platforms. Illustrated SLA terms, including the overall SLA
document are subject to service type and business criteria
with respect to their formulation, machine interpretation and
operational processing.
V. SLA DIGRAPH MODEL
We use the WSLA specification [1] as a reference to
construct a directed graph for the SLA data model. We first
transform primary WSLA language components into element
sets by taking into account their cardinalities and relationships.
Table II illustrates identified element sets. We use a subscript
to denote their cardinality in any SLA graph instance.
SLA {Signatory2, Obligations1,
ServiceInfo1,{ServiceDescriptioni}}, where
i 2 [1, n]
Signatory2 SupportPartyi, where i 2 [0, n]
Obligations ◆ SLOi, ActionGuaranteei, where
i 2 [0, n]
ServiceDefinitioni ◆ ServiceObjecti, where
i 2 [1, n]
ServiceObjecti SLAparameteri, where i 2 [1, n]
CompositeMetrici ◆
ResourceMetrici, CompositeMetricj, Functionj,
where i 2 [1, n] and j 2 [0, n]
i, j indicate the cardinality of element subsets and
n 2 N.
TABLE II: SLA language components - Set representation
The digraph follows the guidelines of the WSLA spec-
ification with respect to language components. We add an
additional node, the service-information one, to include data
that identify overall service properties, which are not related
with obligations of any involved party. Motivation for the
inclusion of such node is derived from the WS-Agreement
specification [3] that contains a similar section named ”Con-
text”. The service-information node denotes SLA attributes

Fig. 3: SLA representation as a digraph
like the location of the provider infrastructure and customer
stored data, the legislation that the service provider adheres
to, as well as service monitoring and configuration options for
customers, if such exist.
Figure 3 illustrates the generated SLA skeleton digraph. The
proposed model consists of 12 nodes that are interconnected
through 19 edges. The assembly of nodes and edges represents
the SLA skeleton with respect to primary SLA language
elements. The ”SLA” represents the source node and the graph
contains in total two sink nodes. The graph model provides
a simple but powerful way to order semi-structured SLA
data into structured context. Moreover, it can be extended
accordingly to depict service specific properties and metrics,
as well as terminology requirements and metric functions of
particular business settings.
By definition, a digraph contains nodes and directed edges,
which in our context represent the SLA content and are
used as indicators of unidirectional dependencies in the
graph information flow. Nodes and edges in the digraph are
accompanied with properties like description, measurement
functions as well as measurement type and unit.
For example, edges that connect SLA parameters to com-
posite or resource metrics, include a property on the measure-
ment and evaluation schedules of the respective metrics. Such
property is needed in order to schedule the value update of
an SLA parameter with respect to the metric values that are
read and used. Such service details are important for SLA
management since they enable the accurate execution of the
service contract. A node or edge property is defined as a key-
value pair that is embedded in the node or edge definition.
Furthermore, in Figure 3 a thick line denotes edges that
indicate strong dependencies between nodes, while a thin line
denotes weak dependencies. A strong dependency signifies
operations between a set of nodes that are connected through
a set of directed edges (e.g. the content of the antecedent node
requires the precedent node for its definition or measurement).
Weak dependencies indicate semantic interrelations between
nodes.
As discussed in Section IV, SLA components have opera-
tional characteristics that require a flexible and dynamic data
structure. For example, SLA parameters and SLOs may need
to be informed at frequent time intervals on the state or value
of other SLA terms. The term unidirectional dependencies is
used to signify operational functions between diverse SLA
components that require an antecedent - precedent relationship
for the evaluation of every pair of SLA terms.
We model this relationship as following: For every pair of
SLA nodes (A, B) that are connected through a directed edge
labeled x: Avalue = function(Bvalue), while a condition is
valid and given that x represents an outgoing edge from A
to B. (A, B) may denote any set of SLA terms. We consider
(A, B) as distinct element sets, except if A represents a sink
node. In such a case, B cannot represent the incoming node
of an outgoing edge from A and respectively A may either
require to apply some function to itself (e.g. update a previous
value, which will inevitably cause a cycle in the graph) or read
and update its value directly from a source (i.e. URI, query
interface). Moreover, we assume that an SLA data operation
is accompanied by a condition that is reflected as an edge
property in the graph and that specifies the operation time
validity, the obliged party for the operation execution and the
required data sources.
For example, in the skeleton graph of Figure 3 a ”SLA
parameter” and a ”CompositeMetric” represent a pair of SLA
nodes, where the ”SLA parameter” requires the value of the
”CompositeMetric” in order to compute its own value. Sim-
ilarly, the ”CompositeMetric” requires the ”ResourceMetric”
value. Yet, the ”ResourceMetric” represents a sink node in the
graph, thus data operations with respect to the retrieval and
update of its value, may either include a cycle (in case an
iterative function is used) or a process may need to read the
metric’s value from an indicated resource endpoint.
The skeleton digraph is implemented as a Python 2.7
module using the NetworkX [7] scientific library. The SLA
skeleton graph is defined as a DiGraph object. The ”Obli-
gations”, ”SignatoryParty” and ”ServiceDescription” nodes
are defined as subgraphs of the SLA skeleton digraph. The
Python module reads lists of ordered service description
details as input and returns SLA service description trees
that follow the set hierarchy of Table II. SLO and action
guarantee descriptions are defined as nodes of the obligations
subgraph. Their dependencies on service description nodes are
represented through edge weights. Generally, SLA component
dependencies are reflected via weights that are assigned to
edges according to data content and operation criticality. The
latter is subject to the SLA instance, service execution time
and SLA application criteria.
For example in the initial Python module implementation,
edge weighting is driven by the update frequency rate of
SLA terms (e.g. measurement of composite, resource metrics).
Moreover, node and edge properties signify specific SLA
conditions (qualitative or quantitative), which in turn affect
the content value of other SLA components or the overall
information flow of an SLA graph instance. For example, given
available service provisioning options in an SLA template,
a customer can select desired parameters to derive a well-
suited service offer. The digraph model eases the automated
identification of available SLA term combinations as informa-

tion nodes and their properties can be retrieved by following
directed paths. For example, a graph traversal can be designed
to return service offer combinations that satisfy customer
requirements, while preserving the provider’s capacity and
availability bounds.
By default, every SLA is associated with a single provider-
customer pair to represent the one-to-one agreement rela-
tionship and thus the ”Signatory party” node has exactly
two instances in the SLA digraph. The party instances are
defined as customer and provider subgraphs and can include
any key-value paired information relevant to each signatory.
Customer and provider can be affiliated with one or more third
parties that support service operations. Both signatories can be
associated with the same third party, following the guidelines
of [1].
The module outputs an SLA digraph that preserves the
directed relationships of all subgraphs and sugraph elements.
The produced digraph can be fed as a JSON dictionary
into a graph-aware database or to any applicable DBMS for
further processing. Alternatively, NetworkX natively supports
the transformation of a graph into its adjacency matrix form.
The skeleton graph can be extended to include any type of
nodes or edges in any valid graph orientation, e.g. diverse
service-description subgraphs with respect to specific applica-
tion or business needs. The digraph representation highlights
the logical flow between service dependencies and guarantees
in the SLA content. Alternative graph representations will
also yield reasonable SLA structures, given business domain
constraints or application specific needs.
VI. USE CASES AND GRAPH MANIPULATION OBJECTIVES
The SLA formalization as a property digraph provides
important advantages. First of all, it allows for data flow
clarity with respect to semantic richness and to hierarchical
ordering of elements. The mapping of SLA information into
nodes and edges enables the storage and management of data
in a structured way. Graph elements can be handled separately
and combined dynamically. Moreover, this formalization en-
ables the classification of information dependencies and their
automatic retrieval through regular path queries (RPQs).
The directed structure helps to view SLA processing op-
erations as workflows that are scheduled to perform data
operations over semi-structured, dynamical data-series (e.g.
value measurements, updates, alert notifications). This is par-
ticularly important when information has to be processed
concurrently and a workflow strategy is needed to handle the
efficient execution of data tasks. The realization of the graph as
directed permits the orchestration of asynchronous operations
that consider dependencies between antecedent and precedent
nodes or node-sets.
Our approach broadens the SLA utilization as it introduces
new manipulation potentials that can generate business value
in the cloud service economy. The SLA digraph model of
Section V is used for our experimental study to help us
evaluate the modularity, query expressiveness and clear infor-
mation flow of the proposed SLA data model. By applying the
model over distributed, virtual hosts and by processing SLA
information with respect to the aforementioned characteristics,
the efficiency of the proposed schema can be examined.
Levels of information granularity may differ within the same
SLA graph as the graph may include subgraphs and traversal
trees for the information assortment. SLA data can be volatile
and their values may influence the values of nearby nodes.
With the term modularity we refer to the structural separation
of SLA components and to their dynamic combination through
graph traversals over multiple node-sets.
Walks within a graph typically take place through regular
expressions that consider both the graph structure as well
as inclusive nodes and edges. According to [8], a property
graph represents a multi-relational graph, where nodes and
edges contain attributes that designate associated key/value
properties. With respect to SLA manipulation, property graphs
can prove valuable as they can encapsulate all the information
required to thoroughly represent an SLA document. We use the
term query expressiveness to denote the easyness (when com-
pared to a human language and to similar quering techniques),
along with the clarity and flexibility to define and execute
query expressions over a graph. Instead of direct regular
expressions, a property-graph domain specific language (DSL)
can be used to interact with underlying graphs.
Due to the primary SLA usage as a measurement instru-
ment, SLA management requires clear and granular informa-
tion flow such that information can be retrieved and processed
rapidly. The structured accommodation of the SLA content
into nodes and edges permits their direct retrieval. Moreover,
a backend DBMS that is graph-aware and thus supportive
of graph operations, can enable the graph management from
the application layer ( e.g. through a RESTful web inter-
face). Generally, the SLA graph manipulation may involve
asynchronous computational tasks to achieve the simultaneous
handling of data updates and condition evaluation. Moreover,
it may require simultaneous HTTP operations with respect to
client-server architectures and data-retrieval requests. Finally,
it allows to study and examine alternative storage and schedul-
ing strategies with respect to combinations of in-memory and
disk data storage.
Thus, we concentrate our current experimentation over
two primary use-cases, where the proposed SLA data model
applies and which can help us collect knowledge on the previ-
ously mentioned graph manipulation objectives. The exercise
of these scenarios allows for the evaluation of the SLA digraph
as an efficient data model for SLA and service management.
Both cases are positioned in the cloud computing setting. For
our experimentaion we assume that information exchange is
primarily processed through the HTTP layer. Moreover, we
consider that computational resources and data repositories are
distributed over remote hosts.
Customer aware SLAs: As we demonstrated in [4], a
virtual marketplace that operates over HTTP, can use SLA
facets as means of service-offer selection features. A customer
can adjust desired service provisioning needs, submit their
requirements to a remote server and receive back offers that

satisfy their requests. A client - server architecture enables
to evaluate the SLA digraph efficiency with respect to scal-
ability and performance throughput over HTTP. The server
keeps SLA offers stored in remote repositories. On customer
request, the server processes one or more queries to backend,
distributed repositories and returns the result to each customer
instance. The scenario involves the parallel processing of
selection and filtering queries with data updates. In [4] we
utilized a relational (MySQL) and a document (MongoDB)
database to experiment. Currently, the proposed graph data
model is deployed for the Titan database [9] and the Gremlin
DSL [10] is used to express the graph data queries.
SLA matching: The SLA Python module of Section V
enables the automated and customized SLA generation. In a
real-business scenario, provider and customer may negotiate
on the desired service levels before agreement initiation. An
SLA can be dynamically formulated by combining resources
and services from a set of providers. Thus, SLA templates
can be generated on-demand. Given provider capacity and
availability constraints that can be modeled as graph node and
edge properties, as well as customer service criteria, the use-
case allows for the simulation of an automated negotiation sce-
nario between two potential signatories. Moreover, a possible
candidate for the formalization and respectively computation
of provider constraints is the Network Simplex algorithm [11].
Simplex represents a network optimization algorithm that
applies smoothly to economic problems, e.g. price setting
optimizations. In the SLA and service management context,
Simplex can be applied to the digraph model through node
and edge properties. For the use-case under consideration, the
algorithm can be applied to examine and model the provider
overall cost from satisfying customer service objectives. Cus-
tomer satisfaction and quality assurance are, undoubtedly,
criteria that any provider needs to consider as their levels
can have consequences to the provider reputation and market
survivability. The NetworkX scientific library provides an in-
hand Simplex implementation, where given a digraph with
weighted nodes and edges, the algorithm returns the minimum
network cost flow that satisfies all node demands.
VII. RELATED WORK
To our knowledge, limited scientific work focuses entirely
on SLA data management. In [12] the authors elaborate on
SLA semantic graphs using an in-depth analysis of service
guarantees and objectives that are derived from real service
offers. The authors define an SLA as a legal contract that
specifies the minimum expectations and obligations between
a provider and a customer. They refer to their model as the
necessary means to design database schemas for the SLA
information. Their SLA semantic model is based on UML and
is formed as a relationship diagram that indicates the logical
information flow among principal SLA elements.
The approach in [12] can be considered complementary
to ours. The authors label their identified graph relationships
with semantic, unidirectional attributes to describe the primary
data flow among business SLA components. Their model
includes the notion of service package graphs and resembles a
hypergraph of service bundles that contains transition triggers
from one service package to another according to predefined
conditions. The SLA model in [12] signifies business perspec-
tives related to service provisioning. Our model proposes a
modular and extensible approach to structure and manage the
SLA information flow.
In [4] we prototyped and experimented with a client -
server scenario, where the server simulates the role of a
cloud service marketplace that uses SLA templates as service
offers. In that scenario, customers submitted their service
provisioning requirements over HTTP and the marketplace,
returned SLA templates that matched customer requests. For
our observations in [4] we deployed a MongoDB and a
MySQL database to store and manage SLA templates. We
measured and reported the overall time of simultaneous query
execution and client - server information exchange over HTTP.
Moreover, we justified that a modular SLA data schema
enhances the data manipulation possibilities.
VIII. CONCLUSIONS AND ON-GOING WORK
The paper discussed SLA data management requirements
by analyzing the SLA anatomy according to the WSLA
specification [1]. SLA language components were introduced
and mapped to data management attributes with respect to
their operational relevance during any SLA activity. The
provided SLA data analysis is a useful tool for the design
and deployment of SLA data stores. Additionally, it is generic
enough to include new SLA elements as well as data attributes
and to adapt into diverse business settings and application
criteria.
Next, we introduced our SLA digraph model that follows
the WSLA language directives and has been developed as
a Python 2.7 module. The programming module uses the
NetworkX scientific library [7] . The module expresses core
SLA components, like service description, obligations and
signatory parties, as subgraphs of the primary SLA digraph.
The digraph module sits on the application layer. It can be
used by backend systems and diverse applications for the
automatic generation of SLA templates. Most importantly, it
can be easily extended to consider conditions and parameter
costs in the SLA formulation.
Although NetworkX provides an excellent library for the
automatic graph generation and manipulation, as it supports
multiple graph algorithms, it unfortunately lacks of a native or
supportive DBMS solution. Thus, as our research concentrates
on SLA data management, we are currently experimenting
with the proposed digraph model using the Titan [9] graph
database and the Gremlin [10] query language.
Our testing is targeted towards the evaluation of the pro-
posed digraph model with respect to the challenges that were
discussed in Section VI. In particular, our current experiments
compare the usage of a graph query language like Gremlin
with XQuery statements over the native WSLA XML format.
Furthermore, the selection of a distributed graph database as

the backend layer enables a real simulation of the first use-
case in Section VI. With respect to the second summarized
use-case our current plan is to re-use the NetworkX library
and possibly adapt the Simplex algorithm to our SLA model
needs.
Summarizing, currently the SLA usage by IT service mar-
kets is superficial and lacks automation and technical details
that would help customers evaluate the service levels that they
are paying for. A graph-based, multi-relational approach for
the SLA formalization and operations creates new opportu-
nities for SLA-aware, value-added services and enhances the
research opportunities around service and SLA management.
ACKNOWLEDGEMENT
This work is supported by the Swiss National Science
Foundation (SNSF), grant number 200021E-136316/1
REFERENCES
[1] H. Ludwig, A. Keller, A. Dan, R. King, and R. Franck, “Web Service
Level Agreement (WSLA) Language Specification,” IBM Corporation,
2003.
[2] A. Dan, H. Ludwig, and G. Pacifici, “Web service differentiation with
service level agreements,” White Paper IBM Corporation, 2003.
[3] A. Andrieux et al., “Web Services Agreement Specification (WS-
Agreement),” Open Grid Forum, (GRAAP) Working Group, 2005.
[Online]. Available: http://www.ogf.org/documents/GFD.192.pdf
[4] K. Stamou, V. Kantere, and J. Morin, “SLA template filtering: a faceted
approach,” in Proc. of the 4th International Conference on Cloud
Computing, GRIDs, and Virtualization, IARIA CLOUD COMPUTING
2013, Valencia, Spain, 2013.
[5] M. Bailey, “The Economics of Virtualization: Moving Toward an
Application-Based Cost Model,” VMware, Nov 2009.
[6] A. Keller, U. Blumenthal, and G. Kar, “Classification and Computation
of Dependencies for Distributed Management,” in Proc. of the Fifth
IEEE Symposium on Computers and Communications (ISCC 2000), ser.
ISCC ’00. IEEE Computer Society, 2000.
[7] A. Hagberg, D. Schult, and P. Swart, “NetworkX,” http://networkx.
github.io/, , accessed: March, 2013.
[8] M. Rodriguez, “Property Graph Algorithms,” http://markorodriguez.
com/2011/02/08/property-graph-algorithms/, accessed: July, 2013.
[9] ThinkAurelius, “Titan Distributed Graph Database,” http://thinkaurelius.
github.io/titan/, accessed: July, 2013.
[10] ThinkAurelius team, “Gremlin graph query language,” https://github.
com/tinkerpop/gremlin/wiki, accessed: July, 2013.
[11] J. B. Orlin, “A polynomial time primal network simplex algorithm
for minimum cost flows,” in Proc. of the seventh annual ACM-SIAM
symposium on Discrete algorithms, ser. SODA ’96, 1996, pp. 474–481.
[12] C. Ward, M. J. Buco, R. N. Chang, and L. Z. Luan, “A Generic SLA
Semantic Model for the Execution Management of e-Business Outsourc-
ing Contracts,” in Proceedings of the Third International Conference on
E-Commerce and Web Technologies, ser. EC-WEB ’02. London, UK:
Springer-Verlag, 2002, pp. 363–376.

SLA data management criteria

More Related Content

What's hot (17)

Similar to SLA data management criteria (20)

Recently uploaded (20)

SLA data management criteria