Modeling for cloud
Semantic Modeling for Cloud Computing
Amit Sheth and Ajith Ranabahu • Wright State University
Cloud computing has lately become the attention grabber in both academia and industry. The promise of seemingly unlimited, readily available utility-type computing has opened many doors previously considered difficult, if not impossible, to open. The cloud- computing landscape, however, is still evolving, and we must overcome many challenges to foster widespread adoption of clouds. The main challenge is interoperability. Numerous vendors have introduced paradigms and services, making the cloud landscape diverse and heterogeneous. Just as in the computer hardware industry’s early days, when each vendor made and marketed its own version of (incompatible) computer equipment, clouds are diverse and vendor-locked. Although many efforts are under way to standardize clouds’ important technical aspects, notably from the US National Institute of Standards and Technology (NIST), consolidation and standardization are still far from reality. In this two-part article, we discuss how a little bit of semantics can help address clouds’ key interoperability and portability issues.
Figure 1 shows the three main flavors of clouds
as outlined by NIST (http://csrc.nist.gov/groups/SNS/cloud-computing). Infrastructure-as-a-
service (IaaS) clouds have the largest gap (a
high workload but little automation) in terms of
deploying and managing an application. Platform-as-a-service (PaaS) or software-as-a-service
(SaaS) clouds have substantially lower workloads,
but at the expense of flexibility and portability.
Given this diverse environment, a cloud service consumer faces four challenges.
First, depending on the application’s requirements, legal issues, and other possible considerations, the consumer must select a cloud to
use. Each cloud vendor exposes these details in
different formats and at different granularity
Second, the consumer must learn about the
vendor’s technical aspects (service interface,
scaling configuration, and so on) and workflow.
Third, the consumer must then develop an
application or customize the vendor-provided
multitenant application to fulfill his or her
requirements. When doing this, the consumer
must take into account various technical details
such as the choice of programming language
and limitations in the application runtime,
which will all be vendor-specific.
Finally, after deploying the application, if
the consumer must change the service provider
(which happens surprisingly often), at least two
major considerations arise. First, the consumer
might need to rewrite or modify the application
code to suit the new provider’s environment. For
some clouds (such as IaaS), this is minimal, but
porting the code in PaaS and SaaS clouds will
likely require more effort.
The second consideration is that data collected for the application might need transformation. Data is the most important asset the
application generates over time and is essential
for continued functioning. The transformation
might even need to carry across different data
models. The industry practice is to address such
To overcome these challenges and provide
better insight into the aspects requiring attention, proper modeling in this space is essential.
Semantic modeling can help with this.
First, a word about cloud interoperability is in order. Interoperability requirements in the cloud landscape in Figure 1 arise owing to two types of heterogeneities.The first is vertical heterogeneity — that is, within a single silo. We can address this by using middleware to homogenize the API and sometimes by enforcing standardization. For example, the Open Virtualization Format (OVF ) is an emerging standard that allows migration of virtual-machine snapshots across IaaS clouds. The second type is horizontal heterogeneity—that is, across silos. Overcoming this is fundamentally more difficult. Each silo provides different abstraction levels and services. High-level modeling pays off, especially when you must move an application and code horizontally across these silos. Surprisingly, many small and medium businesses make horizontal transitions. PaaS clouds offer faster setup for applications, and many exploit the free hosting opportunities of some platform cloud providers (for example, Google’s App Engine). When the application grows in scope and criticality, however, an IaaS cloud might prove cheaper, more flexible, and more reliable, prompting a transition. The following discussion on semantic models applies to both vertical and horizontal interoperability. The key in addressing both types is that many of the core data and services causing them follow the same semantic concepts. For example, almost all IaaS clouds follow conceptually similar workflows when allocating resources, although the actual service implementations and tools differ significantly. Similarly, the PaaS modeling space is a subset of that for IaaS, from a semantic perspective. These observations prompt us to argue for semantic models’ applicability, especially to supplement interoperability in the cloud space. Some parts of the scientific and engineering community weren’t impressed by early semantic-modeling approaches, especially ones that required large up-front investment. This perception, however, is changing rapidly with the influx of new applications and technologies that exploit detailed semantic models to provide improved functionality (for example, biomedical ontologies cataloged at the US National Center for Biomedical Ontology ). Semantic models such as ontologies can formalize more details than other traditional modeling techniques. They also enable reasoning, a way to make inferences and gain new knowledge. Such rich models can improve traditional functions far beyond what was once thought possible. For example, Powerset (now part of Microsoft), a search service provider, added considerable value to search results by incorporating semantic models and thus enabling fact discovery. These capabilities are being complemented by the ability to more rapidly create domain models, often by mining crowd knowledge represented, for example, in Wikipedia1 or shared data, exemplified by the linked-object data cloud. The technologies showcased at the Semantic Technology Conference also provide plenty of evidence of semantic-empowered commercial and scientific applications.
Multidimensional Analysis of Cloud-Modeling Requirements
We suggest a 3D slicing of the modeling requirements along the following dimensions (see Figure 2). The types of semantics that are useful for porting or interoperability in cloud computing are similar to those we introduced in 2003 for Web services. This is natural because the primary means of interacting with a cloud environment is through Web services. The four types of semantics—data, functional, nonfunctional, and system—are based on the different semantic aspects a model must cover.
The language abstraction level indicates the modeling’s granularity and specificity. Although ontological modeling is preferable at a higher level, developers prefer detailed, concrete syntactic representations. These representations of different granularities might need to be related, often through explicit annotations. For example, although service developers happily use Web Services Description Language (WSDL) descriptions in their Web services, these descriptions are syntactic and can’t provide useful semantic details. To overcome this deficiency, SAWSDL (Semantic Annotations for WSDL and XML Schema; www.w3.org/TR/sawsdl) attaches semantic-model details to WSDL documents. The software lifecycle stage is important in determining the modeling requirements. For example, some nonfunctional and system requirements might not be modeled during development but will be taken into account only during deployment. A different team handles each of these lifecycle stages; this separation is important so that one team doesn’t step on another’s toes. This separation aims to focus the modeling effort on the correct time and people. Some cloud models fall under the nonfunctional/system/ ontology space in Figure 2. Such models include the Elastic Computing Modeling Language (ECML), Elastic Deployment Modeling Language (EDML), and Elastic Management Modeling Language (EMML), all based on OWL and published by Elastra (Elastra Languages). However, some aspects of cloud modeling have received little or no attention. For example, there’s no comprehensive higher-level modeling in the data and functional spaces. Lessons learned during large-scale ontological modeling in the Semantic Web and Semantic Web services, biology, and many other domains are readily applicable here and would help address some of the challenges in the cloud space.
Opportunities for Semantic Models in Cloud Computing
Semantic models are helpful in three aspects of cloud computing.The first is functional and nonfunctional definitions. The ability to define application functionality and quality-of-service details in a platform-agnostic manner can immensely benefit the cloud community. This is particularly important for porting application code horizontally—that is, across silos. Lightweight semantics, which we describe in detail later, are particularly applicable. The second aspect is data modeling. A crucial difficulty developers face is porting data horizontally across clouds. For example, moving data from a schema-less data store (such as Google Bigtable1) to a schema-driven data store such as a relational database presents a significant challenge. For a good overview of this concern, see the discussion of customer scenarios in the Cloud Computing User Cases White Paper. The root of this difficulty is the lack of a platform-agnostic data model. Semantic modeling of data to provide a platform-independent data representation would be a major advantage in the cloud space. The third aspect is service description enhancement. Clouds expose their operations via Web services, but these service interfaces differ between vendors. The operations’ semantics, however, are similar. Metadata added through annotations pointing to generic operational models would play a key role in consolidating these APIs and enable interoperability among the heterogeneous cloud environments.
From a perspective of the cloud landscape based on the language abstraction and type of semantics (that is, viewing the cube in Figure 2 from the top), we see that opportunities exist to use semantic models to define applications’ functional aspects in a platform-agnostic manner. In most cases, however, converting a high-level model directly to executable artifacts pollutes both representations. Intermediate representations are important to provide a convenient conversion. Applying high-level modeling to describe an application’s functional aspects isn’t new. Many software development companies have been using UML to model application functionality at a high level and use artifacts derived from these models to drive development. This process is commonly called model-driven development. This is an example of using high-level models to derive fine-grain artifacts. UML models usually don’t include code, so you can use them only to generate a skeletal application. That is, low-level details are deliberately kept away from the high-level models. UML models, however, are inherently bound with object-oriented languages, and UML- driven development processes depend heavily on advanced tools (for example, IBM’s Rational Rose). This limits UML’s applicability. A popular alternative to such tool-dependent heavy upfront models is domain-specific languages (DSLs). Their popularity is due partly to the availability of extensible interpreted programming languages such as Ruby and Python. Unlike UML, a DSL is applicable only in a given domain but enables a light-weight model in that domain, often without requiring proprietary tools. For example, you can use IBM’s Sharable Code DSL (ISC), which is a mashup generator, with a basic text editor. (However, providing graphical abstractions and specialized tooling would be more convenient for users.) “Lightweight” signifies that these models don’t use rich knowledge representation languages and so have limited reasoning capabilities. Our Cirrocumulus project for cloud interoperability (Cirrocumulus) uses DSLs to bridge the gap between executable artifacts and high-level semantic models. A DSL, although domain specific, can provide a more programmer-oriented representation of functional, non-functional, or even data descriptions. A best-of-both-worlds approach is to use annotations to link models, which provides the convenience of lightweight models while supporting high-level operations when required. Figure 2 shows an annotation referring to an ontology from a fictitious DSL script for configuration. The script is more programmer-oriented (in fact, it’s derived from Ruby) but lacks an ontology’s richness. However, the annotation links the relevant components between the different levels, providing a way to facilitate high-level operations while maintaining a simpler representation. From the perspective based on the type of semantics and software lifecycle stage—that is, looking at the cube in Figure 1 from the front — you can see the modeling coverage for software deployment and management. Elastra’s Elastic Computing Modeling Language (ECML),
Elastic Deployment Modeling Language
(EDML), and Elastic Management
Modeling Language (EMML) ontologies cover many system aspects
and some nonfunctional aspects
in all stages. We’re pleased to see
such industry initiative in adopting
Another opportunity for semantic models for clouds lies in Resource Description Framework (RDF) data modeling. As we discussed in part 1, a major concern plaguing cloud computing’s adoption is data lock-in , that is, the inability to port data horizontally. Many vendors designed schema-less, distributed data stores with relaxed consistency models to provide high availability and elasticity to suit clouds’ needs. However,
exploiting these data stores requires substantial redesign of many data-driven applications and often makes porting data to a traditional relational database extremely difficult. The current practice is to address such transitions case-by-case. A better approach is to model the data in RDF and generate the specific target representations, and in some cases even the code for the application’s data access layer. This method can formulate transformations from one representation to another using the lifting-lowering mechanism. Semantic Annotations for WSDL and XML Schema (SAWSDL) demonstrated this mechanism’s use for data mediation.
Lightweight modeling in terms of
DSLs also applies here. For example,
the Web services community has
long used XML Schema definitions
as platform-agnostic data definitions.
Schema definitions serve as inputs to
code generation tools that generate
platform-specific data definitions.
From the perspective of the type
of semantics and software lifecycle
stage, most of this data modeling
applies during application development. Concrete artifacts generated
from these high-level models would
be used mostly during subsequent
One feature differentiating the cloud
from other distributed environments
is the availability of Web services
to manipulate resources. Availability of the service APIs lets you
programmatically manage the cloud
resources, even from within the
same cloud. These capabilities have
revolutionized application deployment and management. For example,
you can compose or mash up well-defined services to facilitate elaborate workflows.
Service definitions are usually
syntactic, and many researchers
have focused on embedding rich
metadata in formal service descriptions. One result of this research
is SAWSDL. A growing trend is
to annotate HTML descriptions to
embed richer, machine-readable
semantic metadata. One reason for
this method’s popularity is search
engines’ use of metadata to display results in customized formats.
Yahoo’s SearchMonkey and Google
Rich Snippets are two such microformat-driven schemes. These annotations, unlike the DSL annotations
in Figure 3, might not always point
to ontologies. Their structure can be
based on a vocabulary or taxonomy—
a lower-grade nonsemantic model.
For example, the popular hCalendar microformat is part of the “lowercase
semantic web” movement, which
emphasizes lightweight models.
Embedding rich semantic meta-data in cloud service descriptions has
three main benefits that go beyond
customized search capabilities.
The first benefit deals with Representational State Transfer (REST)
style services. Many cloud service
providers adopt REST-style Web services that don’t advocate a formal
service description. These services
are described using HTML pages.
WSDL 2.0, the latest specification,
explicitly supports formal description of “RESTful” services but hasn’t
seen quick adoption. Alternative
approaches such as SA-REST(SA
stands for semantic annotation), a
generic annotation scheme that follows microformat design principles,
are becoming more applicable in
this space. These annotations enable
the seamless, flexible integration of
formalizations into RESTful service
descriptions. This opens the door
to many exciting avenues such as
faceted search to identify relevant
reusable services and semiautomated
The second benefit deals with
handling change. The cloud space is
still evolving. If the history of software or component interoperability is any guide, achieving consensus in
the cloud space will be difficult and
won’t likely happen soon. Attaching
formalizations via annotations, however, is flexible enough to accommodate an evolving model. This is
especially attractive to vendors who
aren’t willing to invest heavily in
The third benefit is that the formalizations apply not only to service descriptions but also to many
other aspects such as service level
agreements (SLAs) and software
licenses. You can use annotations
to embed formalizations even for
these documents, facilitating more
automation in the cloud space. For
example, Web Service Level Agreement (WSLA) specification provides a
way to formalize SLAs, but creating
and maintaining these formalizations is timeconsuming.
Figure 3 illustrates using SA-REST annotations on the Amazon
Elastic Compute Cloud (EC2) SLA
document. It shows how a capable
processor could use these annotations to extract a WSLA equivalent
of the human readable SLA.
These benefits’ importance comes
into perspective when you consider
the enormous body of research on
standard-driven service compositions and agreement matching.
The informal, non-standard-driven
nature of many cloud services made
most of the previous research inapplicable. However, being able to
glean formalizations from existing
documents opens the doors to apply
many well-researched techniques.
The cloud space presents many opportunities for researchers, and we see a plethora of applications that use semantic modeling. Issues such as interoperability and data portability, which the cloud community is facing right now, are the very issues for which semantic models excel in providing solutions. However, learning from the past, we advocate a multilevel modeling strategy to provide smooth transitions into different granularity levels. We also think that DSLs can play an important role in the cloud space to provide lightweight modeling in an appealing manner to the software engineering community.
1. C. Thomas et al., “Growing Fields of Interest—Using an Expand and Reduce Strategy for Domain Model Extraction,”Proc. 2008 Int’l Conf. Web Intelligence and Intelligent Agent Technology (WI-IAT 08), vol. 1, IEEE CS Press, 2008, pp. 496–502.
2. K. Sivashanmugam et al., “Adding Semantics to Web Services Standards,” Proc.Int’lConf.WebServices (ICWS 03), CSREA Press, 2003, pp. 395–401.
3. F. Chang et al., “Bigtable: A Distributed Storage System for Structured Data,” in Proc. Usenix Symp. Operating Systems Design and Implementation, Usenix Assoc., 2006, p. 15.
4. M. Nagarajan et al., “Semantic Interoper- ability of Web Services—Challenges and Experiences,” Proc. 2006 IEEE Int’l Conf. Web Services (ICWS 06), IEEE CS Press, 2006, p. 373–382.
5. A.P. Sheth, K. Gomadam, and J. Lathem, “SA-REST: Semantically Interoperable and Easier-to-Use Services and Mashups,” IEEE Internet Computing, vol. 11, no. 6, 2007, pp. 91–94.
Amit Sheth is the director of the Ohio Center of Excellence on Knowledge-Enabled Computing (Kno.e.sis) at Wright State University. He’s also the university’s LexisNexis Ohio Eminent Scholar. He’s on the Web at http://knoesis.org/amit/.
Ajith Ranabahu is pursuing a PhD in cloud- computing interoperability at Wright State University. He worked with IBM on Sharable Code and its Altocumulus project, and he coordinates the Cirrocumulus project (Cirrocumulus). Contact him at firstname.lastname@example.org