Sunteți pe pagina 1din 6

Towards improving resource discovery in Grid computing using semantic and multi-agent

methodology

Abdul Khalique Shaikh1, Saadat M. Alhashmi2, Rajendran Parthiban3

School of IT, Monash University, Sunway Campus, Malaysia,


1
shaikh@infotech.monash.edu.my
2
saadat.m.alhashmi@infotech.monash.edu.my
3
School of Engineering, Monash University, Bandar Sunway, Malaysia,

rajendran.parthiban@eng.monash.edu.my

Abstract
=====================================================================

Grid computing can be used to aggregate physical and logical resources for executing compute-intensive
applications. However, due to the heterogeneous and dynamic nature of computational Grid systems, efficient
resource discovery is a very challenging issue. Agents can be used to enhance resource discovery processes, since
they are distributed in nature and autonomous and intelligent in behavior. Furthermore, agents can interact with
neighboring agents based on local knowledge and move along predetermined routes autonomously on behalf of a
user. Ontology can provide benefits in a resource discovery mechanism. These benefits include enhancement of
interoperability, quality of information, accuracy and efficiency. In this paper, we propose a comprehensive solution
for a decentralized resource discovery process by improving the ontological framework. This proposed work is
based on extending the combined semantic and agent approaches. The aim of this research is to extend and develop
the decentralized agent and semantic-based resource discovery process to improve the performance and maximize
the utilization of resources in a computational Grid. Preliminary simulations carried out using the proposed approach
show more tasks being accepted for complex jobs. Consequently, this approach could be useful in increasing the
utilization of resources in a Grid Computing environment.

Keywords: Grid Computing, Resource Discovery, Semantic Technology, Multi-agent system, Ontology
is not scalable and causes poor performance in a large
1. Introduction scale Grid system. On the other hand, a distributed
Grid computing is a distributed computing mechanism is more successful than centralized
environment aimed at sharing geographically because it possesses characteristics of scalability,
distributed resources [1]. Resource discovery is a dynamism, load balancing, and heterogeneity and
function of Resource Management System (RMS) fault resilience. However, a syntactical distributed
which is a central part of Grid computing and approach is based on exact keyword matching and
responsible for collaboration of these distributed provides limited results. To overcome these issues,
resources in an efficient way [2]. As the resources in semantic technology has been introduced. In the
Grid computing environment are dispersed and semantic approach, semantic metadata and ontology
dynamic in nature, the maximal utilization and are used to describe the resources which are very
collaboration of these resources can be challenging. helpful to discover and monitor Grid computing
Furthermore, these resources are managed resources. Agents with semantic technology can
independently by various organizations with their improve the resource discovery process as they are
own local policies [3]. Such kind of behaviors of a distributed in nature and can move to predetermined
Grid system affects the resource discovery process routes autonomously. By integrating agent and
radically. semantic-based approach, the resource discovery
process in Grid computing is significantly improved.
Centralized resource discovery approach is We propose a comprehensive solution for a
preferred in small and medium size Grids which decentralized resource discovery process by
provides easier control and management. However, it enhancing the ontological framework for which, we
develop sub-domain specific ontologies for each Grid searching mechanisms rely on string matching [1].
resource. We also use agents for avoiding overhead This mechanism is based on fixed query method
of negotiation where resource agents need to know which gives very limited results. A DHT (Distributed
neighbor connections and semantic descriptions of Hash Table) based search algorithm is introduced in
neighboring resource agents. Based on local [10] where the author has designed and implemented
knowledge, the resource agent interacts semantically SWORD, a Scalable Resource Discovery service for
with neighboring agents and forms resource chain for Wide-area distributed system. Author has claimed
a task. This proposed work is based on extending the that DHTs are important distribution application
integration of semantic and agent approaches. building blocks because a service built on a top of
DHT automatically inherits the DHT‟s self
The remaining sections of this paper are configuration, self healing and scalability. However
organized as follows: the paper [6] mentions that SWORD provides
network coordinates as a location information but the
Section 2 describes reviews and outlines latency is not sufficient for band-width demanding
limitations in existing resource discovery systems. applications. The focus of SWORD is only to provide
Section 3 explains multi-agent and semantic scalability in terms of querying [11]. In all the above
technologies. Section 4 presents implementation and syntactical distributed approaches, it is possible to
results. Finally section 5 provides conclusion and miss relatively capable resources because it is based
possible future work. on exact key words matching and current schedulers
in Grid Computing do not possess the capability to
infer the relationship among resources [12]. To
2. Related work overcome these issues, a semantic approach has been
Significant research has been done to introduced.
improve resources discovery component of Grid A peer-to-peer semantic-based system Atlas
computing in both academia and industry. This proposed in [13] for distributed storage and retrieval
section discusses some of the state of the art in this of RDFs data. This is built on top of the DHT
domain before proposing how it can further be bamboo and supports pull and push query scenario. It
improved. is being used currently to realize the metadata service
of S-OGSA (Semantic Open Grid Service
Condor [4] uses centralized queries for Architecture) in a fully distributed and scalable way.
resource discovery which works on the principle of a Authors claim that resource discovery services for
matchmaking approach [5]. The matchmaking is semantic grid can be made scalable, fault tolerant,
based on an idea that resource providing services and robust and adaptive. However, the process of query
clients requesting service advertise their evolution, measurement and experimental results are
characteristics and requirements using classified missing.
advertisements (ClassAds). The matchmaking A novel semantic supported agent-based
framework includes several components of a resource decentralized Grid resource discovery mechanism has
discovery mechanism which provides stateless been proposed in [14]. This heuristic algorithm has
matching services [6]. MDS [7] is a suite of been developed to find out the neighbors resources
components for monitoring and discovering resources and introducing semantic similarity of domain
and services which offered by Globus. It uses ontology using the decentralized approach. The
standard interfaces defined within the WS-Resource experimental results show that the success probability
Framework (WSRF) and WS-Notification (WS-N) of resource discovery increases with decrease in the
specifications to provide query and subscription similarity threshold. The level of the task complexity
interfaces. Both Condor and MDS follow centralized is inversely proportional to success probability.
and hierarchical approaches framework respectively Authors claim that the algorithm has a flexibility to
and work fine for a small to a medium scale Grid. efficiently and dynamically discover resources.
However, for large-scale Grids, these are not scalable However, their experiment results show that the task
and cause poor performance. Also it poses single success probability is very low under average task
point of failure [8], and do not guarantee scalability complexity and average semantic threshold values. In
and fault tolerance [9]. this experiment authors use a general domain specific
ontology with artificial generated similarity threshold
On the other hand, distributed approach in a values. However, in reality, each type of resource is
resource discovery process is very helpful in terms of different in terms of function. This is the reason we
self-organization, adaption, load balancing and fault use sub-domain specific ontologies for each resource
tolerance but it has a number of limitations such as separately.
To make a comprehensive resource similarity of concepts represents the degree of
discovery model, we propose extending the combined commonality between concepts. The authors of paper
agent and semantic approach for accurate and [20] derive conceptual similarity using the notion of
autonomous resource discovery. The semantic “similarity graph” which is a subpart of the ontology
matching is very flexible and it can enhance represented as a graph with subset of concepts as
interpretability in a Grid environment [15]. Agents nodes and relations connecting these as edges. Also
are extremely helpful to provide methodologies for authors of this paper derive a function sim (x,y) that
consuming, processing, creating and disseminating of measures the degree of similarity proportional to how
information and knowledge within a distributed much concept x and y share or how close they are.
framework. These objectives could be achieved by We compute semantic similarity based on the
improving ontological frameworks using sub-domain following formula which is published in the same
specific ontology for each type of resources paper.
separately. Semantic threshold values must be
computed by an appropriate formula. Equation 1

where is a factor that determines the degree of


3. Multi-agent and Semantic influence of generalization of ontology concepts. The
Technology value of „ ‟ lies in between 0 and 1. If value of ‟ ‟ is
This section presents features of a Multi- 1 that means perfect generalization where each and
agent system and a semantic technology. It also every concept is defined and 0 means very poor
explains how these can be supportive in resource generalization. We use the value of P = 2/3 (0.67)
discovery process in Grid computing. which is used to compute the semantic similarity
threshold in paper [20] with the same degree of
3.1 Multi-agent Technology generalization, is the set of nodes (upwards)
Agents can be useful for analysis, design reachable from x and is the reachable
and realization of a distributed environment such as nodes shared by x and y. The lowest values
Grid. Agents have ability to manage dynamic means not similar and the highest
properties [16, 17]. They provide better solutions in value of means fully similar i.e
terms of adding autonomy, adaptability and control
intelligence to the Grid services [18]. To search for We use four types of resources i.e Processor,
appropriate resources, agents can move by Memory, Storage and Operating System in our
predetermined routes autonomously and process a job experiment. To compute semantic similarity, we use
in a very flexible way. An individual agent can query two existing ontologies for resource „Processor‟ and
neighbor nodes without need of global knowledge „Operating System‟ from paper [21] and developed
and can semantically interact with neighbor agent- ontologies for other two types of resources i.e
based on local knowledge [14]. In our experiment „Memory‟ and „Storage‟. The Ontology for
there are 64 agents on provider side and four types of „Memory‟ resources is as follows:
resources are uniformly distributed among these
agents. Each agent has its own label i.e. A1,Q2, Ab2
etc. Each agent is randomly connected with some
other agents who are called neighbors of that agent.
Those neighbor agents keep information about other
agents which play an important role in making
resource chain for a task.

3.2 Semantic Technology


The performance of resource discovery
depends on the quality of information. Semantic
technology provides meta information which
Figure 1: Ontology for ‘Memory’ Resource
describes data semantic and supports interoperability
[19]. Semantic similarity threshold values can be With the help of four sub-domain specific
calculated with the help of ontologies of Grid ontologies and Equation 1, we compute the values of
resources. semantic similarity threshold between concepts of
3.2.1 Semantic Similarity: resources. The semantic similarity threshold values of
Semantic similarity is defined as how „Memory‟ resource concepts are as follows:
concepts of ontology are related to each other. The
4.1 Searching Mechanism
Suppose a user selects two tasks and set the
task complexity value 0.25 with the semantic
similarity threshold 0.4, it means each task requires 4
resources to get the task done with semantic
similarity equal or greater than 0.4 values. We
assume that task 1 requires one „Processor‟, One
„Memory‟, One „Storage‟, and one „Operating
system‟. However, name of resources for this task 1
are randomly generated from ontology concepts for
respective resource types i.e „Processor32‟ „RAM‟
„HardDisk‟ and „MacOs‟ respectively. The task 2
requires three „Memory‟ and one „Operating system‟.
To make a complete service chain of required
resources for these two tasks separately, searching
mechanism compiles the following steps.
First of all, search algorithm works for the
Table 1: Semantic Similarity Threshold Values task 1 required resources i.e „Processor32‟ „RAM‟
of ‘Memory’ Resource „HardDisk‟ and „MacOs‟. Search algorithm selects
an agent „K1‟ randomly who has resource
3.2.2 Task Complexity
„Processor‟ now this agent compares its resource
Task complexity is measured on
with one of the required resources i.e „Processor32‟
the basis of length of the tasks which means
„RAM‟ „HardDisk‟ and „MacOs‟ based on semantic
resources required by tasks are different. If task
similarity threshold. In this case, the similarity
requires maximum number of resources then the task
threshold between „Processor‟ and „Processor32‟ is
complexity is high and if task requires minimum
0.78 and user has set 0.4 threshold it means
resources then the task complexity is low. We define
semantic similarity of the agent is greater than the
the length of tasks 4, 8, 12 and 16 for each task. If
required resource, then algorithm checks if
task complexity is set 0.25 it means one task needs 4
neighbors of K1 that hold resources i.e (Processor,
resources and 0.5 means 8 resources, 0.75 means 12
Storage, Memory, OS) has at least one of the
resources and 1 means one task requires 16 resources.
required resource other than randomly picked agent
resource i.e (RAM, HardDisk and MacOs ) and their
status should be free. In this case it meets above
4. Implementation conditions and randomly picked node K1 who has
We extend the decentralized resource agent- resource „Processor‟ is committed for the task 1.
based semantic resource discovery algorithm used in In case of the semantic similarity threshold
paper [14]. To implement resource discovery model, is not matched, this step repeats until and unless we
we opted to work on Netlogo simulation and get the right resource. In 2nd step, randomly
modeling tool. selected another node i.e „AP3‟ who has resource
We evaluate the performance of the „Storage‟, algorithm requires the agent AP3
resource discovery model in terms of task success
compares its own resource i.e „Storage‟ with one of
probability. We observe the task success probability
the required resources i.e „RAM‟ „HardDisk‟ and
under different network and under different
„MacOs‟ on basis of semantic threshold. In this case
semantic threshold. Our objective is to extend this
work to make resource discovery accurate and semantic similarity threshold between „Storage‟ and
autonomous by improving ontological framework. „Harddisk‟ is 0.78 which is greater than pre-set
We use sub-domain specific ontology for each type value of semantic threshold. It means condition is
of resource separately. In this implementation, we matched, then checks if neighbor of AP3 node
use realistic semantic similarity threshold values resources i.e OS, Memory Storage and Processor
which are calculated using Equation 1. We have has at least one resource which is already
created 64 agents on provider side and uniformly committed, in this case „Processor‟ is matched
distributed four types of resources among these because it is already committed. Now randomly
agents. picked node which has resource „Storage‟ are
committed for the task 1. At this stage we have two
committed resources i.e Processor, HardDisk and
two required resources RAM and MacOs. Repeats
randomly selecting one another node Bj4 who has
resource „OS‟ first it compares the semantic
similarity threshold with one of required resource, in
this case semantic similarity threshold between OS
and MacOS is 0.78 which is greater than required
semantic similarity threshold then checks neighbors
of BJ4 has resources „Processor‟ „Storage‟ „OS‟
„Memory‟, one of these resource node is already
committed with the task 1 so resource „OS‟ is
committed on behalf of „MacOS‟, for last resource
of the task 1, randomly selecting one another node
„Ad2‟ who has resource „Memory‟, the algorithm
requires compares this resource with one of the Figure 3: Result of simulations under task
required resource i.e. „RAM‟ on basis of semantic complexity 0.50
similarity threshold. The value is 0.78 between
„Memory‟ and „RAM‟ which is greater than pre set Above graphs show the relationship between
value of semantic similarity threshold, then checks the task success probability and the semantic
if neighbors of Ad2 who has resources (Storage similarity under different levels of the task
Processor and OS) and one of these should be complexity and the different number of tasks. The
committed for the task 1 then finally the randomly results show that the task success probability is
picked node who has resource „OS‟ is committed. higher when using low semantic similarity threshold
Now the resource chain for the task 1 has completed and the success probability is lower when high task
and it is ready to work with required resources. complexity is set. The results show that the greater
Subsequently 2nd task, once again, follow the number of tasks affects the low task success
similar steps and so on. If any task doesn‟t fulfill the probability. We have compared our results with paper
algorithm steps then it is rejected and algorithm [14] and found our results better than that. The reason
works for next job. of getting better task success probability is that we
have used sub domain of specific ontology for each
type of resource and we have computed semantic
5. Results threshold values by appropriate formula. However
We have run 200 simulations with different under 0.8 and 1.0 semantic threshold our results show
values of the task complexity, the number of tasks 0 task success probability. The reason is, a
and the semantic similarity threshold. The results of classification widens the similarity gap. Hence, when
these simulations are as follows: the semantic values are increased, the task success
probability in our experiment is almost 0.
Our overall results indicate that the extending the
integration of semantic and agent approach by
enhancing ontological framework can be useful in
increasing the utilization of resources and improving
the performance of resource discovery process which
is beneficial for both providers and users.

6. Conclusion and Future Directions


It can be concluded that a resource searching
mechanism in a Grid Computing environment must
be flexible and effective so that tasks can be run with
Figure 2: Result of simulations under task appropriate required resources in a given timeframe.
complexity 0.25 An extension of a combined agent and semantic-
based decentralized resource discovery model has
been presented and implemented in this paper. The
proposed model allows individual resource agents to
find neighboring agents-based on local knowledge
rather than on global information. We have used sub-
domain specific ontologies for each resource Generation Computer Systems, vol. 23, no. 7,
separately and computed semantic threshold values 2007, pp. 864-878.
on the basis of ontological concepts. We compared 10. J. Albrecht, et al., “Design and implementation
our results with existing work and found our results trade-offs for wide-area resource discovery,”
to be slightly better in terms of task success ACM Trans. Internet Technol., vol. 8, no. 4,
probability. Our experimental findings show that the 2008, pp. 1-44.
success probability of resource discovery increases 11. M. Cardosa and A. Chandra, “Resource Bundles:
with the decrease in similarity threshold. The level of Using Aggregation for Statistical Wide-Area
the task complexity is inversely proportional to the Resource Discovery and Allocation,” Proc. The
task success probability. As for future work, we are 28th International Conference on Distributed
planning to generate detailed ontologies for each grid Computing Systems IEEE Computer Society,
computing resource which include synonyms, 2008.
hyponyms and hypernyms. Subsequently we will 12. B.R. Amarnath, et al., “Ontology-based Grid
extend this resource discovery model into large scale resource management,” Software: Practice and
to check the scalability of the system. Experience, vol. 39, no. 17, 2009, pp. 1419-
1438.
References 13. Z. Kaoudi, et al., “Semantic Grid Resource
1. J. Li, “Grid resource discovery based on Discovery in Atlas,” Knowledge and Data
semantically linked virtual organizations,” Management in GRIDs, 2007, pp. 185-199.
Future Generation Computer Systems, vol. 26, 14. H. Liangxiu and D. Berry, “Semantic-Supported
no. 3, 2009, pp. 361-373. and Agent-Based Decentralized Grid Resource
2. M. Beltrán, et al., “Resource Dissemination Discovery,” Future Generation Computer
Policies on Grids,” On the Move to Meaningful Systems, vol. 24, no. 8, 2008, pp. 806-812.
Internet Systems 2004: OTM 2004 Workshops, 15. L. Chen and F. Tao, “An Intelligent
2004, pp. 135-143. Recommender System for Web Resource
3. A. Caminero, et al., “Extending GridSim with an Discovery and Selection,” Intelligent Decision
architecture for failure detection,” Proc. Parallel and Policy Making Support Systems, 2008, pp.
and Distributed Systems, 2007 International 113-140.
Conference on, 2007, pp. 1-8. 16. Z. Li and M. Parashar, “Rudder: An agent-based
4. K. Krauter, et. al., “A taxonomy and survey of infrastructure for autonomic composition of grid
grid resource management systems for applications,” Multiagent and Grid Systems, vol.
distributed computing,” Software: Practice and 1, no. 3, 2005, pp. 183-195.
Experience, vol. 32, no. 2, 2002, pp. 135-164. 17. Y. Jian, et al., “Grid Resource Management
5. A. Sharma and S. Bawa, “An Improved Based on Mobile Agent,” Proc. International
Resource Discovery Approach Using P2P Model Conference on Computational Inteligence for
for Condor: A Grid Middleware,” Proc. World Modelling Control and Automation and
Academy of Science, Engineering And International Conference on Intelligent Agents
Technology, 2006. Web Technologies and International Commerce
6. J. Kim, et al., “Using Data Accessibility for (CIMCA'06), 2006, pp. 255-255.
Resource Selection in Large-Scale Distributed 18. J.W. Yin, et al., “Weaving an agent-based
Systems,” IEEE Transactions on Parallel and Semantic Grid for distributed collaborative
Distribution System, vol. 20, no. 6, 2009, pp. manufacturing,” International Journal of
788-801. Production Research, vol. 48, no. 7, 2009, pp.
7. J. Schopf, et al., “Monitoring and discovery in a 2109 - 2126.
web services framework: Functionality and 19. D. Fensel, Ontologies: a silver bullet for
performance of the globus toolkit's mds4,” knowledge management and electronic
Argonne National Laboratory, Preprint commerce, Springer Verlag, 2004.
ANL/MCS-P1248-0405, 2005. 20. T. Andreasen, et al., “From ontology over
8. Hauswirth and R. M. Schmidt, “An overlay similarity to query evaluation,” Proc. 2nd
network for resource discovery in Grids,” Proc. CologNET-ElsNET Symposium-Questions and
Sixteenth International Workshop on Database Answer: Theoretical and Applied Perspective,
and Expert Systems Applications, 2005, pp. 343 2003, pp. 39-50.
- 348 21. A.C.T. Vidal, et al., “Applying semantics to
9. P. Trunfio, et al., “Peer-to-Peer resource grid middleware,” Concurr. Comput. : Pract.
discovery in Grids: Models and systems,” Future Exper., vol. 21, no. 13, 2009, pp. 1725-1741.

S-ar putea să vă placă și