Abstract: This study discusses the design and simulation of an e-learning computer network topology based on Grid computing technology for Indonesian schools called the Indonesian Education Grid (abbreviated as IndoEdu-Grid). The usage of such a grid has its advantages such as supporting the use of heterogeneous resources (hardware or software) and resources sharing. The establishment of such network without Grid computing capabilities will lead to redundancies of the idle resources. The simulation which is built using GridSim toolkit will handle the proposed two conditions or scenarios that have different network topologies based on their routers and links configuration. Each scenario will be run in the simulator using two packet scheduling algorithms, one will be FIFO (First In First Out) Scheduler and the other SCFQ (Self-Clocked Fair Queuing) Scheduler. The processing time of the job’s packets will be evaluated to determine the most effective network topology for IndoEdu-Grid. The simulation result showed that if SCFQ scheduling algorithm is used then the most appropriate network topology is the topology which allows its packets with the same priorities to have less possibility to collide one another in one router or link which is the first scenario. The simulation results also showed that SCFQ scheduling algorithm can reduce the packets lifetime at routers that have very crowded traffics. This fact implies to the decrease of the whole job processing time.

INTRODUCTION

E-learning as a trend has been developed so rapidly that many educational organizations and institutions in numerous countries including Indonesia adopts it. An example of an infrastructure that can be used for e-learning in Indonesia is Jardiknas (Jejaring Pendidikan Nasional or Indonesian ICT Network). Jardiknas is a national-scaled WAN (Wide Area Network) that facilitates educational activities in Indonesia. This network consists of Institutional/Official Jardiknas that serves online data transaction between educational institutions, College Jardiknas-Indonesian Higher Education Networks (INHERENT) that serves science and technology research and development, School Jardiknas that serves information and e-learning accesses in schools and Teacher and Student Jardiknas that serves personal information and e-learning accesses (Pustekkom, 2009). Jardiknas is established with a main purpose to serve administration in central National Department of Education (Departemen Pendidikan Nasional) and many domestic or foreign related work units and to serves learning processes in primary, junior high and senior high schools based on information and communication technology. Jardiknas covers the whole areas of Indonesia but it still has some problems as follows:

To solve these problems, we propose the usage of Grid computing technology for Indonesian Education Networks called IndoEdu-Grid. Grid computing (or Grid) is a system that can provide resources sharing among organizations. Grid infrastructure will provide a capability to connect the resources dynamically as an ensemble to support large-scale, resource-intensive and distributed applications (Berman et al., 2003).

IndoEdu-Grid itself is a national-scaled Grid network used for educational purposes in Indonesia especially for high schools and universities. This network consists of ten resources with many processors located in each province that connect to each other. However, building a suitable and reliable national Grid infrastructure is very expensive, thus this encourages us to study some possible topologies and perform some experiments on their performances.

It is impossible to perform such an experiment using the trial and error method because it will be very expensive and time-consuming. In addition, we have to consider both political and technical policies are considered in each province in Indonesia.

Thus, prior to the implementation, the experiments must be designed, done and analyzed using a simulator that can be configured and adjusted to THE needs. This simulator has to mimic the real system and predict system’s behavior. In this study we use GridSim toolkit which is a Java-based simulator and supported with some additional libraries.

GridSim toolkit (or GridSim) is a toolkit developed with Java and is used to build discrete-event simulations of Grid systems. GridSim is an open-source application and licensed under GPL license, thus it encloses its source codes in its distribution package. GridSim’s rationale is that creating a testbed infrastructure for Grid system is expensive and time-consuming, even an existing testbed infrastructure is also limited in size to a few resources and domains and testing scheduling algorithms for scalability and adaptability and evaluating scheduler performance for various applications and resource scenarios is harder and impossible to trace (Buyya and Murshed, 2000, 2002; Sulistio et al., 2005).

MATERIALS AND METHODS

We design the IndoEdu-Grid, it consists of entities such as resources, users and jobs or Gridlets; class diagram and topological scenarios such as:

Resources: Resource entities are responsible to perform computation on job entities in form of Gridlets sent by one or more users and send it back to the user.

The research uses one resource for each province; each resource consists of one Machine and each Machine consists of 4 PEs (processing elements). These Machines have identical processing speed (300 MIPS), operating systems and architectures (Debian Linux on Intel architecture).

To reduce the level of heterogeneity and make the understanding of system behavior easier, all provinces will use the same resource characteristics. The details of name and location of resources and the amount of each PE for the entire province are shown in Table 1.

Table 1:	Resources for each Province

Users: Users are entities responsible to submit jobs in form of Gridlet objects to the resources. In this research, the users are programmed to send jobs to a particular resource at the same time, thus we are able to gain more knowledge on the performance of Grid system in its peak load when all the users are accessing the resource at the same time.

The selection of resource is done randomly, for example the users located in Banten may choose the resource located in Lampung and send jobs to that resource.

The number of users in each province is adjusted by the real conditions according to the number of high school students taken from the data published by National Department of Education. The most recent data published by the National Department of Education is the data in 2007/2008 academic year.

Therefore, the research used the number of high school students in 2007/2008 academic year. The number of users are adjusted to fit with the simulation. This adjustment is performed by finding the actual ratio of the number of students in each province to the number of students in Special Capital District of Jakarta. This adjustment is made to prevent the occurrence of OutOfMemory runtime error in Java. The real total number of all students is 15, 112, 161 so we think that this adjustment is important. The ratio is then applied to the simulation so that the number of users in each province can be obtained. The obtained number of users after adjusted is 958.

Jobs (Gridlets): Jobs in GridSim are represented as the objects of the class Gridlet provided by GridSim. In the research, each user will create three Gridlets having different lengths-5000 MI (millions instructions), 3000 MI and 1000 MI. This was aimed to simulate the real situation where a user does not just send one job but it can also send more than one job with different sizes and needs of computation powers.

A user will send the three Gridlets successively to a resource that has been chosen randomly. Furthermore, the resource will send an ACK (acknowledge) signal as a sign that the Gridlet been received and ready to be processed. Resources will process the Gridlet and send it back to the user.

The class diagram: The class diagram is shown in Fig. 1. This simulation consists of four main classes, namely class Main, class Islands, class NetUser and class Randomizer and several other supporting classes available in Java's default library and the GridSim’s additional library such as ArrayList, Router, GridSim, GridResource and GridletList. The function of the classes are as follows:Main class is the class containing the Main method that will run the simulation and the printGridletList method that will print the processed Gridlets. In this class, the construction of the network topology for each scenario will be done with the following stages:

Fig. 1:	The class diagram for simulation program

Islands class is a class that contains methods for creating GridResource objects. In this class, there are two methods, they are createGridResourcesList and createGridResource. NetUser class is a class that represents the users in Grid system. In this class, the behavior of the objects of user in the Grid system is set through the body method. Randomizer class is a class that functions to do randomization between two integers. This class has only one method, Randomize method which has two integer parameters which serve as initial boundary and final boundary for the random numbers.

GridSim class is a class responsible for initializing, running and stopping the whole process of simulation. In this class, there are important methods such as init to initialize GridSim, startGridSimulation to run simulations and stopGridSimulation to stop the simulation process. In addition there are also the methods for logging the statistics of simulation obtaining the status of Gridlets, stopping the simulation process for a while (pausing the simulation), sending (submitting) the Gridlets to specific resources and obtaining the ID of the entities. GridletList class is a class responsible for storing some Gridlets into a list. This class extends LinkedList class.

GridResource class is a class responsible for simulating resources in Grid systems. The properties of GridResource are defined in the objects from ResourceCharacteristics class. Router class is a class that simulates routers on the network.

This class is extended by some classes that implement different routing method such as FloodingRouter, FlowRouter, FnbRIPRouter, RIPRouter and RateControlledRouter classes. RIPRouter class is a class that implements the use of RIP (Routing Information Protocol).

This class is actually a part of the GridSim library but some parts are adjusted to the needs of this simulation in order to be able to write the routing table of routers to files named [router_name].out. PacketScheduler class is an interface that provides a template for the schedulers that will be used on the router.

In the research, this class is implemented by two scheduler classes, namely FIFOScheduler and SCFQScheduler. FIFOScheduler and SCFQScheduler classes as the names imply, respectively implement FIFO and SCFQ scheduling algorithms to schedule packets at routers. ArrayList class responsible in storing objects into a list.

Fig. 2:	Three types of router linking

This class represents a standard data structure that is owned by the Java programming language.

The topology scenarios: The simulation consists of two scenarios divided based on the configurations of the router and link connections. We use two different types of scenarios to see whether the effects of the formation or hierarchy of links and router configurations which connect the provinces on one island and then connect and divide these islands based on the Western, Central and Eastern Indonesian zone have significant impacts on the performance of Grid systems. The two configurations are based on the thought that the territory of Indonesia is composed by three components or areas the western parts of Indonesia, the central parts of Indonesia and the eastern parts of Indonesia. These three major areas can then be viewed also as entities composed by several islands and/or archipelagos. These islands and/or archipelagos can be viewed as entities composed by one or several provinces. The concept of this dividing will be used in the router configuration for each scenario. The scenarios can be described as follows:

Scenario 1: This scenario is a representation of the proposal that divides the whole territory of Indonesia into three main sections the western, central and eastern part of Indonesia.

Each of these three sections will be subdivided into parts or units that are smaller the islands and/or archipelagos. These units of islands and/or archipelagos are subdivided again into the smallest units namely the provinces. Thus, this scenario will create a multilevel (hierarchical) dividing which starts from three main units (western, central and eastern part of Indonesia). These main units will consist of seven islands and/or archipelagos units (Sumatra Island, Java Island, Kalimantan/Borneo Island, Sulawesi Island, Bali Island- Western Lesser Sundas-Eastern Lesser Sundas, Moluccas Archipelago and Papua Island). In the end, these seven islands and/or archipelagos units will consist of 31 province units (ten provinces in Sumatra Island, six provinces in Java Island, four provinces in Kalimantan/Borneo Island, 6 Provinces in Sulawesi Island, three provinces in Bali Island-Western Lesser Sundas-Eastern Lesser Sundas, one province in Moluccas Archipelago and one province in Papua Island). Figure 3 shows the network topology for the first scenario. The points between the leaf routers symbolize the meaning of until.

Scenario 2: This scenario is a representation of the proposal that divides the whole territory of Indonesia directly into islands and/or archipelagos units. These islands and/or archipelagos will be divided again into province units. This dividing mechanism will create seven islands and/or archipelagos units and 31 province units. Each islands and/or archipelagos unit will consist of some provinces units. Figure 4 shows the network topology for the second scenario. The dash points between the leaf routers means until.

Assumptions in the Simulation: There are some assumptions during the simulation and these are as follows:

Fig. 3:	Network topology for the first scenario

Fig. 4:	Network topology for the second scenario

The computing environment: The simulation is run under Intel^® Core™ 2 Duo T5800 processor with 2.0 GHz clock speed, 800 MHz FSB (Front Side Bus) and 2 MB L2 cache 2048 MB RAM (Random Access Memory) with shared dynamically with Mobile Intel^® Graphics Media Accelerator 4500MHD and 320 GB Fujitsu MHZ2320BH G2 SATA harddisk with 5400 rpm rotation speed. While the software are 32-bit Microsoft Windows Vista™ Business operating system. JDK (Java Development Kit) version 1.6.0_05 with Java™ Runtime Environment 1.6.0_05-b13. GridSim version 5.0 beta.

RESULTS AND DISCUSSION

Simulation results: In the simulation, the -Xmx300 m parameter is used to set the Java heap memory usage to 300 MB. The use of heap memory of 300 MB is done to ensure the simulation running smoothly without the occurrence of OutOfMemoryError runtime exception. In addition, the cp parameter followed by the gridsim.jar file is used again so that JVM can find where the required classes are located.

Table 2:	Average simulation results data for the entire Provinces per Gridlet using FIFO and SCFQ scheduling algorithm

Table 3:	Average processing time for the entire provinces and Gridlets using FIFO and SCFQ scheduling algorithm

Main method is located in Main file, so that Main.class will be executed.

The research compares the simulation results of the average processing time of the three types of Gridlets sent to the resource through three network topologies. The simulation will show which topology that produces the lowest and highest average processing time using FIFO and SCFQ algorithm for scheduling packets. The simulation was run 10 times in each scenario to increase the validity of simulation results and then the results were averaged.

The simulation results data per Gridlet which is the average of all provinces using the FIFO and SCFQ algorithm are shown in Table 2 while the average of all data is shown in Table 3. From Table 3, find that Scenario 1 with SCFQ algorithm gives the best performance (180.30313 simulation seconds) while Scenario 2 with FIFO gives the worst performance (183.20544 simulation seconds).

Analysis: In Scenario 1 with FIFO algorithm, it appears that processing time is large enough, i.e., 183.03943 simulation seconds. This processing time is 0.09% lower compared to the processing time of Scenario 2 using the same algorithm, i.e., 183.20544 simulation seconds. This happened because the load distribution was not done evenly. The large number of users and resources, i.e., 238 users (24.84% of total users) and 10 resources (32.26% of total resources) in Sumatra Leaf Router and 505 users (52.71% of total users) and 6 resources (19.35% of total resources) in Java Router Leaf make data traffic loads in those areas become very large. This also happened to other routers in the western region, such as WestInd Core Router, Java Edge Router and Sumatra Edge Router. In addition, the packets were served without any priorities so that all packets had to wait a long time especially those that came from and go to resources in Java. Eventually, bottlenecks occured.

In Scenario 1 with SCFQ algorithm, it appears that the processing time is 180.30313 simulation seconds or 1.49% lower compared to the processing time when using FIFO algorithm. This processing time is the lowest processing time, so that this scenario is the most appropriate scenario for IndoEdu-Grid. This happened because the load distribution was done more evenly. As mentioned in the previous paragraph, the packet traffics were large enough in the western routers.

However, arrived packets have different priorities, indicated by their weight; packets with weight 1 (high priority packets) will be served first while the packets with weight 0 (normal priority packets) will be served after the high priority packets. This differentiated service will not make the packets wait for a long time and reduce the possibility of the occurrence of bottlenecks.

The fact in Scenario 2 with FIFO algorithm can be explained with the same reasons in Scenario 1, i.e., the uneven load distribution where the routers’ loads in the western part are greater than the other routers and the lack of packets priority made the packets get the same treatment.

The possibility of bottlenecks is very large. However, Scenario 2 has a difference than Scenario 1 in terms of the existence of core routers. The configuration of Scenario 2 does not involve core routers, so the packets with far destination will have greater number of hops. By observing Fig. 3 and 4, this fact can be explained by the following examples. If a user in Lampung wants to send jobs to a resource in Papua, job packets will be propagated through channels as follows:

Table 4:	The packet lifetime in routers using SCFQ algorithm relative to that using FIFO algorithm

The number of hops in Scenario 1 example is 6 hops, while in Scenario 2 example is 7 hops. The maximum amount of hops in this configuration is 7 hops. Thus, the inexistence of core routers will increase the number of hops for the packets with far destination. This is why the processing time of Scenario 2 with FIFO algorithm is higher than Scenario 1 with the same algorithm and makes this scenario to be the worst performance scenario Table 4.

We also computed packet lifetime the difference between the enqueuing and dequeuing time of packets in routers using SCFQ algorithm relative to the ones using FIFO algorithm in all scenarios. The value less than one indicates that the lifetime in routers using SCFQ algorithm is less than that in the routers using FIFO algorithm. In general, SCFQ algorithm causes lower packet lifetime, except only in three routers that are in CentralInd Core Router (in Scenario 1) and Sumatera Edge Router (in Scenario 2). This happens because the use of SCFQ algorithm caused overhead time to select the packets to be processed. These three routers and the other routers with value close to one are connected with high-speed link (100 Mbps, 1 Gbps and 500 Mbps), so that the packets through these routers can be transmitted directly without having to wait for long time in the queue. This also eliminates such routers need to the use of SCFQ algorithm.

CONCLUSION

In this study, IndoEdu-Grid network using GridSim toolkit is stimulating. The simulation was conducted with two different types of topologies in terms of link and router configurations and two types of scheduling algorithms FIFO and SCFQ. The average processing time is studied to obtain the most effective topology for each scheduling algorithm and the final results of the packet lifetime is used to analyze the phenomenon in the simulation.

If the FIFO scheduling algorithm will be used in establishing IndoEdu-Grid network, then the most appropriate topology is the topology that allows the packets to have a low number of hops. In this simulation, the network with the lowest number of hops is provided by the third scenario. The use of topology in the first and second scenario only makes the processing time becomes longer because the packets will have a greater number of hops.

If the SCFQ scheduling algorithm will be used in establishing IndoEdu-Grid network, then the most appropriate topology is the topology which makes the data packets with the same priorities have little chances to meet each other in a single router or link. In this simulation, topology that meets these conditions is the topology on the first scenario. Topology on the second does not meet these conditions, so it is the not appropriate topology for SCFQ scheduling algorithm.

SCFQ scheduling algorithm tends to make packet lifetime in routers with crowded traffic becomes shorter. Packet lifetime shows the difference between the enqueuing and dequeuing time of packets. This is because there are packet priority settings where the packets with higher priority will be served first, so the overall packet lifetime will be reduced.

RECOMMENDATIONS

Based on simulation results in the research, we recommend that the network topology in the first scenario with the implementation of SCFQ scheduling algorithm is used as a reference for establishing IndoEdu-Grid. The first scenario with the implementation of SCFQ scheduling algorithm has the highest level of effectiveness in terms of job packets propagation, although the achieved level effectiveness is very small (<1%). This is because this research used the size and number of Gridlets that are not too large (5000 MI, 3000 MI and 1000 MI), so the savings or effectiveness only slightly visible.

ACKNOWLEDGEMENT

We thank Muhammad Rifki Shihab for reading the first draft of this study and suggesting important improvement.