Distributed computing power service coordination based on peer-to-peer grids architecture

(1)

Distributed computing power service coordination based

on peer-to-peer grids architecture

Min-Jen Tsai

*

_{, Yin-Kai Hung}

Institute of Information Management, National Chiao Tung University, 1001 Ta-Hsueh Road, Hsin-Chu 300, Taiwan, ROC

Abstract

As computational speed, storage capacity, and communication technologies steadily progress to the advanced levels, many business employments have utilized these improvements to aggregate distributed resources in various enterprise collaborating functionalities. Recently, grid computing has been identified as a critical technology by industry for enterprise computing and business-to-business com-puting. The concept of peer-to-peer (P2P) network is also evolving to an expanded usage in distributed networks for sharing the resources like the content files or real-time data for enterprise applications. Therefore, it is natural to include the grids and P2P to support environments that have features of both limiting case. In this paper, a new computing design – NaradaBrokering based Computing Power Services (NB-CPS) – has been applied to utilize the P2P grid to integrate the computational grids, distributed objects, P2P net-works under the hybrid environment. In this study, the approach has been applied to analyze the robustness of digital watermark by filter bank selection and the performance can be improved in the aspect of speedup, stability and processing time. NB-CPS is not only suitable for executing computing works which are able to run in batches, but also be able to solve current issues in Web Services based Com-puting Power Service (WS-CPS) such as system resilience, fault tolerance, efficiency of job scheduling and the instability in congested network environment.

Keywords: Distributed computing; Grid; NaradaBrokering; Peer-to-peer; Web services

1. Introduction

As computational speed, storage capacity, and commu-nication technologies steadily advance, increasingly large, complex, and resource-intensive applications are being developed both in research institutions and commercial industry. The power of network, storage, and computing resources has important implication for successful opera-tion to the academy research and enterprise business ( Casa-nova, 2002. Recently, grid computing has been identiﬁed as a critical technology by industry for enterprise computing and business-to-business computing (Foster, Kesselman, & Tuecke, 2001). The improvements in wide-area network-ing make it possible to aggregate distributed resources in

various collaborating institutions and to form what have come to be known as computational grids (Foster, 2002). To date, most grid applications have been in the area of sci-entiﬁc computing as scientists world-wide are resorting to numerical simulations and data analysis techniques to investigate increasingly large and complex problems.

Due to the cost considerations, small and medium-sized enterprises (SME) generally like to apply the P2P comput-ing which is defined as the sharcomput-ing of computer resource and services by exchanging the information and content files, processing cycles, cache storage, and disk storage for data files. Through sharing the computation load and the computing power by the existing desktop and network, it allows the enterprise to reduce the cost and efficiently leverage their total computation powers and benefit the entire business. Comparing to P2P and grid, both are for sharing of varied resources, ranging from programs, files, and computation capabilities to computers, sensors, and

*

Corresponding author. Tel.: +886 3 571 2121x57406; fax: +886 3 572 3792.

E-mail address:[email protected](M.-J. Tsai).

www.elsevier.com/locate/eswa Expert Systems with Applications 36 (2009) 3101–3118

Expert Systems with Applications

(2)

networks in diﬀerent property and characteristic with ben-eﬁts, it is natural to include the grids and P2P to support hybrid environments that have features of both limiting case.

Based on previous studies, we have proposed the Web Services based Computing Power Service System (WS-CPS)Tsai et al., 2006to fulﬁll SME distributed computing needs. However, there are still some issues in WS-CPS sys-tem needed to be further investigated such as: (1) The requirement of ﬂexibility and fault tolerance mechanism in system architecture. (2) In high-throughput environ-ment, there is a need of network bandwidth control for quality service. (3) The priority mechanism for multi-task demand situation. (4) The improvement of task allocation and re-allocation strategy.

To address these problems, we purpose a novel design named NaradaBrokering based Computing Power Ser-vice System (NB-CPS) which is using NaradaBrokering as the distributed computing middleware ( NaradaBroker-ing Project, 2006). NB-CPS provides a lightweight dis-tributed computing environment to gather computing powers in SME and is suitable for executing computing works in batch format. In addition, it also provides the visual interface to design and monitor the computing tasks.

The proposed design has been applied to analyze the robustness of digital watermark by filter bank selection (Tsai, 2004) and there is not only a significant upgrade in operational efficiency, but also effective task allocation and exception handling mechanisms in multi-tasks with different priorities situation.

This paper will be organized as following: Section2will brieﬂy explain the distributed architectures of WS-CPS and the issues need to be solved. In Section3, solutions for WS-CPS’ problems will be recommended and the proposed NB-CPS will be explained in Section 4. The implementa-tion and performance comparison of NB-CPS and WS-CPS is analyzed in Section5 and conclusion will be given in Section6.

2. Issues with WS-CPS

This section will brieﬂy introduce WS-CPS and some issues should be further studied for improvement.

2.1. The roles of WS-CPS and the operational procedures WS-CPS is bases on the architecture of web services, it inherits the characteristics of Service Oriented Architecture (SOA) which consists of three participants that are service requester, service provider and service broker. The descrip-tion of three roles is as following:

1. The role of coordinator: The coordinator acts as a service broker to fairly mediate between the computing unit (service requester) and computing requester (service pro-vider). Its major function is to maintain a list which

records the URL and requirement of computing ter. This list will be created when the computing reques-ter publishes its Web service in the coordinator. If computing unit asks for the subtasks through the coor-dinator, the coordinator will assign the URL of comput-ing requester in the list to computcomput-ing unit by round-robin mechanism. Afterwards, the computing unit will use the speciﬁed URL to communicate with the comput-ing requester directly. In addition, the function of account and auditing management will be implemented at the end of coordinator. This role is corresponding to the role of UDDI in SOA.

2. The role of computing power requester: The requesters should design their experiment processes and publish their requirements with the coordinator. In addition, it will assign the subtasks to the computing unit with the work ﬂow control capability.

3. The role of computing unit: This role is responsible for executing computation. It will inquire the coordinator for the job while it is idle. After getting back the ter’s URL of Web services, it negotiates with the reques-ter to download the subtasks along with the required data. When the subtask is ﬁnished, it will respond the result to the requestor and the whole procedure will con-tinue until all subtasks are completed.

The operational procedures of WS-CPS are shown in Fig. 1with the following steps:

Step 1: Requestor uses BPEL to design the entire process. Step 2: Requestor registers its task to coordinator. Step 3: Coordinator conﬁrms the request.

Step 4: Computing unit registers its computing power to coordinator.

Step 5: Coordinator sends requestors’ addresses to com-puting units.

Step 6: Computing Unit makes the contract to the requestor.

Step 7: Requestor transfers the tasks to computing units. Step 8: Computing unit sends back the result after

execution.

Step 9: Requestor combines all results while the job is completed.

2.2. The issues of WS-CPS

There are some issues which are needed to be improved for WS-CPS as following:

1. The requirement of ﬂexibility and fault tolerance mecha-nism in system architecture. From the systemic point of view, every node in the peer-to-peer distributed comput-ing infrastructure should have equivalent important position. In other word, every role should be able to be played on any node in the network. However, nodes in WS-CPS communicating with each other through the web service and managing the process by BPEL make

(3)

roles ﬁxed in the system. Therefore, the whole system would be stopped for operation if the role of coordina-tor is not functioning. Therefore, the backup or mirror sites should be able to take over the down machines

and continue the operation when the above mentioned situation happens for the distributed computing system. Certain ﬂexibility and fault tolerance mechanism is required. WEB SERVICE WEB SERVICE BPEL SERVICE WEB SERVICE COORDINATOR COMPUTING UNIT 1: Design workflow 2: Register tasks 7: Transfer tasks 9: Combine results 3: Confirm requests 5: Send addresses 4: Register computing power

6: Contract with requestors 8: Return results 2 1 4 WEB SERVICE BPEL SERVICE BPEL SERVICE 5 3 9 REQUESTOR 6 7 8

Fig. 1. The ﬂow diagram of WS-CPS system.

(4)

2. In high-throughput environment, there is a need of network bandwidth control for quality service. Fig. 2 is the plot which shows the execution time for small and large file transmission in WS-CPS. While the transmis-sion file is small, the transmistransmis-sion time is decreasing when the computing units is increasing. However, the transmission time and the computing units are increas-ing when the transmission file is as large as 130 MB. Due to the constraints of the network bandwidth, there is in practical a limit for the transmitted file size. On the other hand, the number of computers increasing causes the whole network congested for file transmission. Therefore, a proper network bandwidth control for quality service is required for stability consideration. 3. The priority mechanism for multi-task demand situation.

In WS-CPS, there is no design for diﬀerent priority ser-vice which results no diﬀerentiation in serser-vice requests. In fact, there usually exist urgent requests for enterprise and the request of priority service is necessary for qual-ity service and VIP level customers.

4. The improvement of task allocation and re-allocation strategy. While the task is allocated for the computing units, the overall time to get the task done and receive the reply is actually composed by two independent time component: the transmission time and the actual pro-cessing time. During the data transmission period, the computing unit should keep processing other sub-tasks without wasting time for another job if the task can be parallel processed. Therefore, it will be the optimal solu-tion by allocating all subtasks to all idle computing units (Plank et al., 2001) and keeping all units busy till the whole task is done. Under such condition, certain job queue and pre-fetch scheme should be applied to keep the computing units busy.

On the other hand, the distributed computing network is a dynamic environment. The actual processing results could be diﬀered from the forecasting ones after the initial task allocation and certain adjustment is necessary. For example, there are eight subtasks for three computing units as allocated in Fig. 3and apparently PC1 performs better among others. While the sub-task 7 is completed in PC1, the overall job completing time will be T2 instead T3 if the sub-task 8 is re-allocated from PC3 to PC1. A good prediction mechanism is needed for the job re-allocation strategy.

3. Proposed solutions for WS-CPS

To solve the issues mentioned in the previous section, there are proposed solutions for the new design called NB-CPS. The detailed information will be explained as following.

3.1. The application of NaradaBrokering middleware to build ﬂexible and fault tolerant computing sharing environment

To make the system with flexibility and fault tolerance, the entire messaging infrastructure should be re-designed. In this study, we apply NaradaBrokering as the messaging infrastructure to substitute web service structure in WS-CPS and propose the NB-WS-CPS design. NaradaBrokering is a Java based open source distributed messaging infra-structure purposed by Indiana University Community Grids Lab., Publish/Subscribe Model (Pub/Sub)Banavar et al., 1999; Segall and Arnold, 1997; Viktor et al., 2003; is the main communication structure in NaradaBrokering. It provides two closely related capabilities. First, it pro-vides a message-oriented middleware (MoM) that facili-tates communications between entities through the exchange of messages. Secondly, it provides a notification framework by efficiently routing messages from the origi-nators to only the registered consumers of the message in question (Fox et al., 2003). As shown inFig. 4, NaradaB-rokering provides a JMS standard API for developers to create distributed applications, and furnishes an intermedi-ary layer for message transmission.

To oﬀer the characteristics of scaling, load balancing and failure resiliency system, NaradaBrokering is imple-mented on a network of cooperating brokers under the hierarchical structure, where a broker is part of a cluster that is part of a super-cluster which can be demonstrated in Fig. 5. Therefore, the distributed broker network can scale to support the increase in these aggregated entities and the addition of brokers aids the scaling which will not degrade performance by increasing communication

1 4 6 7 2 5 3 8 PC 1 PC 2 PC 3 1 4 6 7 2 5 3 8 PC 1 PC 2 PC3 8 T3 T1 T2

Fig. 3. The task re-allocation process in an ideal environment.

Application A

Application B

JMS Compliant API

Message Oriented Middleware/ NaradaBrokering Network Publish Message Subscribe Message Brokers

(5)

pathlengths or ineffective bandwidth utilizations between broker nodes within the system (Fox et al., 2003). Besides, the disseminations pertain to routing content, queries, invocations etc. to the relevant destinations in an efficient manner. The routing engine at each broker ensures the paths traversed within the broker network to reach destina-tions are along efficient paths that eschew failed broker nodes. The messaging infrastructure is able to communi-cate across firewall, DHCP and NAT boundaries. It also supports various types of protocol such as TCP/IP, UDP, RTP, RMI, HTTP and SOAP.

To use NaradaBrokering as the distributed computing network middleware, it can resolve the flexibility and fail-ure resiliency issues. From the system point of view, the nodes can play different roles by registering different topic and easily substitute the broken nodes to continue the sys-tem functionality in NB-CPS. However, such flexibility and adjustment is not possible for web service based architecture.

3.2. The bandwidth ﬂow control mechanism by FCFS reservation strategy

As noted in the previous section, WS-CPS is instable and will have high tendency to failure due to the network

congestion. To resolve this issue, we design a ﬂow control mechanism as shown inFig. 6 which likes a network gate embedded within the requestor while the transmitting ﬁle is large. The gate control will allow the transmission with the permitted reservation and the reservation queue inside the requester is based on the FCFS strategy as the dis-patching policy. In Fig. 6, the communication between the computing unit B and the requester is allowed since it has the permitted reservation while computing unit A is waiting for the permission since its reservation is still in the queue.

3.3. Weighted priority based task selection

NB-CPS uses weighted priority strategy (Erradi et al., 2006) to select the requested tasks for the scheduler which means the requested tasks with higher priority will have higher chance to be selected than the ones with lower pri-ority. For example inFig. 7, the task from requestor ‘‘A” with priority ‘‘3” will have 3 times higher probability than the task from requestor ‘‘B” with priority ‘‘1”. The unsigned tasks will also have the priority waiting apprecia-tion scheme to avoid the dispatching starvaapprecia-tion.

3.4. Use K-means prediction process to resolve the task allocation and reallocation strategy

In a distributed computing environment, the actual exe-cution time will be different in different computing units. Even executing the same task by the same computing unit, the execution time will be different due to the dynamic sys-tem loading during the execution period. After all tasks are allocated, it is evident that some of the computing units will finish the jobs earlier and some are not. If we can predict the remaining execution time of unfinished tasks correctly, it will be quite natural to reallocate the unfinished tasks from a slow computing unit to a fast one whose expected completion time will be shorter than the remaining execu-tion time at the slow computing unit. However, such a predication process requires the historical data with trust-ful estimation within a dynamic environment. A depend-able scheme of allocation and reallocation strategy is required for NB-CPS.

Rationally, the execution time will be longer if the com-puting unit is in a high loading level and a task during the execution will practically rely more heavily on a speciﬁc Broker Node Service Provider End Client Super-super cluster Super cluster Super cluster Super cluster

Fig. 5. NaradaBrokering network.

REQUESTOR COMPUTING UNIT - B COMPUTING UNIT - A GATE _QUEUE Resource files Result files

Fig. 6. The network bandwidth control strategy.

SCHEDULER

REQUESTOR - A REQUESTOR - B Finished Tasks

Finished Task

Priority = 3 Priority = 1

(6)

system performance indicator named ‘‘critical performance indicator”. For example, the critical performance indicator of Linpack (Linpack, 2006) is CPU usage. Tasks with fre-quent file transmission such as Napster, their critical per-formance indicator will be the network bandwidth. Summing up the related works from Wang et al. (2007), the author analyzed the historical executing data and used value function to assign different weights to different sys-tem performance indicators to further find out the loading level of a computing unit. The authors in Tsai and Wang (2008)dynamically adjusted system performance indicators

by using CC-FGDM for WS-CPS. Those methods gener-ally analyze the historical data of the whole system in order to give the same weighting of the performance indicator for whole system. However, such an approach ignores the dif-ference between each participating party. For example, a computer of high CUP speed with high loading level may have a longer execution time than a computer of slower CPU speed with low loading level. While analyzing the his-torical data of these two machines, the weightings of the performance indicator of CPU speed and loading level should be diﬀerent for each individual computing unit.

Fig. 8. (a) The execution time under diﬀerent loading level for four computing units. (b) The execution time under two loading levels for one computing unit. The red dot is the mean value in each loading level. (For interpretation of the references in colour in this ﬁgure legend, the reader is referred to the web version of this article.)

(7)

FromFig. 8a, it shows the historical data of the execu-tion time and the CPU usage percentage for a distributed computing system with four computers. It is hard to differ-entiate those data and find a general pattern for each com-puter. However, a low loading and high loading pattern for only one computer is easily distinguished inFig. 8b. To fur-ther assist the classification for the loading, support vector machine based classification (Chang & ‘‘Lin, 2001) is applied here for accurate clustering. Therefore, we adopt K-means clustering algorithm (MacQueen, 1967) for NB-CPS since it can be used to group the historical executing data which includes values of critical performance indica-tor and the execution time. The formula is in Eq.(1), where k is the number of clustering from SVM calculation, Si is the ith group, Xjis the jth event value in Si, uiis the mean value for group Si

V ¼X k i¼1 X Xj2Si jXj lij ð1Þ

The system critical performance indicator is obtained in NB-CPS by comparing the correlation values of the pro-cessing time versus the usage. The formula of correlation is in Eq.(2), where X is the indicator usage, X is the aver-age usaver-age, Y is the processing time and Y is the average processing time q¼ P_{ðX X ÞðY Y Þ} ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi P ðX X Þ2 q ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi_P ðY Y Þ2 q ð2Þ

If the correlation value between execution time and sys-tem indicator value greater than 0.7, the syssys-tem indicator will be claimed as the critical performance indicator for the task. Using such a characteristic, we can analyze each computer unit’s behavior and the executing pattern from our veriﬁcation shows two clear clusters which represent high and low system loading levels for CPU usage from SVM classiﬁcation. Therefore, the execution time for low and high loading pattern will be used for task allocation strategy.

The necessary condition of K-means prediction process is to gather enough historical execution information for good estimation. Therefore, the prediction will be more accurate if the computing unit is involved in the distributed computing system for a long enough time period.

4. Proposed NB-CPS design

To sum up the proposed solutions in the previous sec-tion, we purpose the NaradaBrokering based computing power service system (NB-CPS) in this study. NB-CPS is a distributed computing platform based on NaradaBroker-ing (NaradaBrokering Project, 2006) based on P2P grids and provides a lightweight distributed computing environ-ment for SMEs. The comparison between NB-CPS and WS-CPS will be explained in Section 4.1and the detailed

architecture, design and operational processes will be intro-duced in Sections4.2–4.4, respectively.

4.1. The comparison of message transmission infrastructure in NB-CPS and WS-CPS

WS-CPS is a web service based computing power shar-ing platform, all the data transmissions are communicated through the standards of WSDL, SOAP, XML and UDDI. Such infrastructure not only improves the reuse of services but also reduces the ease-to-use and flexibility of system. The requirement of setting up web service server and UDDI increases the entry barrier and causes system roles fixed in specific nodes. Therefore, a role cannot be easily replaced when it crashes and the role regenerating pro-cesses is hard due to the environment restriction. On the other hand, the transmissions of NB-CPS between nodes are reached by Pub/Sub model and nodes in NB-CPS sys-tem play different roles through subscribing to different templates. While certain role disappears in the system, dif-ferent node can use the Pub/Sub model to claim the substi-tution of the missing node and continue the normal function easily.

In NB-CPS system, every event is triggered by node-state change or message receiving. As shown in Fig. 9, a message structure is composed by three parts: ‘‘header” which records routing information, ‘‘template” which describes the subject and the type of the message, and ‘‘payload” which contains the content of the message. Every node must connect to a broker by a unique identiﬁ-cation number. Fig. 9shows a message that a computing unit (id: 1111) asks the requestor (id: 2222) for tasks with the event description of ‘‘AskForTask”. This computing unit sends its performance indicators to the requestor along with the same message.

Since all messages must pass through brokers in NB-CPS, the cost of transmission and the system performance will be decreasing if large transmission files are transmitting across nodes. Thus, messages are categorized into two dif-ferent types of ‘‘event message” and ‘‘resource files”. As shown inFig. 10, the transmission between nodes is peer-to-peer when a message is notified as ‘‘resource files” type to lower the transmission costs in NB-CPS.

Header Template: To/2222 Payload

<Event>

<SenderEntityId> 1111 </SenderEntityId > <Describtion> AskForTask </Describtion> <CPUSpeed> 3120 </CPUSpeed> <CPUUsage> 0.3 </CPUUsage> <RAMUsage> 0.5 </RAMUsage> </Event>

(8)

4.2. System architecture of NB-CPS

There are fours roles in NB-CPS architecture: requestor, scheduler, computing unit and broker as shown inFig. 11. The system functionality can be categorized into four lay-ers: user layer, computing power sharing layer, monitor layer and communication layer. The enterprise employee can design and monitor their tasks at the user layer, con-duct task scheduling, allocating execution and outcome reply in the computing power sharing layer. Each partici-pating node could monitor other nodes’ operation in mon-itor layer and the communication among nodes takes place

in the communication layer. Detailed procedures for each role in each layer will be explained next.

4.3. Roles in NB-CPS

Any distributed computing system usually contains one or more nodes which have the computing requirements (referred to as requestor in NB-CPS), and one or more nodes that can share their computing resources (referred to as computing unit in NB-CPS). There should be a role of resource allocator to deal with complex transactions between these two roles (referred to as scheduler in NB-CPS). In addition, there are independent brokers which are responsible for the transmission of all messages to ensure the successful communication. Under such a P2P grid environment, a node can be a requestor, computing unit or scheduler at the same time. Any node leaves or dis-appears from the system will be monitored by broker and replaced immediately. Roles of NB-CPS are as following:

Requestor

Requestor is the entry point for external users. Users deﬁne task’s quantity, resource, priority. . .etc. to initiate the entire request process. Functions of requestor in each layer are introduced next.

g s g s

s SCHEDULER

BROKER

REQUESTOR COMPUTING UNIT

Event Messa e Event Messa e Resource File

Fig. 10. Two diﬀerent message transfer modes in NB-CPS.

Ci P Ci ti User Layer omput ng owe r Sharing Layer Monitor Layer BROKER

Template Registration Message Routing

ommun ca on Layer REQUESTOR Template Register Message Builder / Parser

Monitoring Process Monitor Scheduler Substitute Scheduler Task Allocation Report Task Transmit Task Collect Result Task Defining Meta-data One-off Resource Constant Resource SCHEDULER Monitoring Process Monitor Requestors

Monitor Computing Units

Requestor Allocation Requestor Assigner Weighted-Priority Based Task Re-allocator K-Means Prediction Requestors Information Computing Units Information Template Register

Message Builder / Parser

COMPUTING UNIT Monitoring Process Monitor Tasks Monitor Status Pre-fetching Process Sending Result Execute Tasks Receive Tasks Process Information QoS Information Template Register

Message Builder / Parser

(9)

User layer: Users design their tasks in the user layer. Since NB-CPS is suitable for executing computing works which are able to be executed in batch format, this kind of task will contain three different types of resources as ‘‘meta-data”, ‘‘one-off resource” and ‘‘constant resource” as shown in Table 1. ‘‘Meta-data” contain the detailed information associate with the task like task name, prior-ity, quantity and broker IP,. . .etc. Resources used by every subtask are named ‘‘constant resource” and used only once per subtask are named ‘‘one-off resource”. Therefore, con-stant resources can be transmitted at the beginning of the process and one-off resources can be transmitted along with each assigned subtask to reduce the transmission cost. Computing power sharing layer: Requestors should elab-orate their computing requirements to the system sched-uler. After computing units assigned by scheduler, requester will transmit the task required information to computing units directly. In addition, requestor will get the associated information about the computing unit so it will analyze the computing unit performance continuously. Monitor layer: Requestors periodically send testing mes-sages to ensure the scheduler is still alive. If the scheduler does not respond to the requester within a given time, NB-CPS will consider that the scheduler is missing and will be replaced by the requestor who finds this event first. The replace process will be detailed later.

Communication layer: Requestors broadcast messages to brokers in communication layer in order to discover and connect to the system scheduler thereafter.

Scheduler

To increase eﬃciency and reduce complexity in the dis-tributed computing system, a resource allocation role is necessary while there are multiple requestors and comput-ing units at the same time. Multiple schedulers are possible for NB-CPS while the network system is growing. Cur-rently NB-CPS uses the ‘‘Centralized Scheduling Model” which coordinates all resources by a single scheduler to increase eﬃciency of the system due to the scale constraint of the network system in SME.

When the ﬁrst requestor submits its request in NB-CPS, it will be assigned as a scheduler in the same time until the node leaves the system. This means there will be many requestors but only the ﬁrst requestor will be the scheduler. Therefore, the scheduler is not there alone in the system but coexists with a requestor. The purpose of such design is applying the P2P characteristic that each node plays equal importance and every node could be the requestor,

computing unit and be the scheduler as well. Since all the requestors will continuously communicate with the sched-uler, the ﬁrst requestor who ﬁnds the scheduler missing will be elected as the scheduler in NB-CPS. Functions of the scheduler are detailed as follows:

User layer: NB-CPS provides a visualized user interface to monitor working progresses which include tasks’ infor-mation, task completed percentage and states of computing units.

Computing power sharing layer: Scheduler is responsible for task allocation. It allocates requestors’ tasks to avail-able computing units according to the task priority. NB-CPS uses ‘‘Weighted Priority Based Task Allocation Strat-egy” (Erradi et al., 2006) and tasks with higher priority will be allocated ﬁrst. After all tasks are allocated, scheduler will start the task re-allocation process to utilize computing powers.

Monitor layer: To ensure all requestors are still alive, scheduler will monitor the requestors by recording the timestamp of triggering messages sent by requestors. If the interval of two trigging messages larger then the pre-deﬁned threshold, NB-CPS will consider that the requestor is already gone, and its remaining tasks will not be allo-cated any more.

Communication layer: Scheduler discovers and connects to requestors and computing units through the broadcast-ing messages by brokers in the communication layer.

Computing unit

The computing unit is the role which can share its com-puting resource in NB-CPS. Once the registered comcom-puting unit is idle, it will send its state and request for job

execu-Table 1

The computing task execution table

Meta-data One-off resources Constant resources Broker IP Executable files Constant files Task’s name One-off files

Priority Result ﬁles Quantity Connection Establishment Scheduler Initialization Task Allocation Scheduler Missing Computing Unit Missing Requestor Missing Broker Missing Task Re-allocation

(10)

tion to the scheduler. The scheduler will then assign an enrolled requestor to the computing unit so both parties can start to communicate for task execution. Functions of computing units are detailed as follows:

User layer: NB-CPS provides a visualized interface for users to monitor system conditions and working progress of the computing unit. The computing unit can decide to stop sharing its computing resource at any time and it can join or leave NB-CPS at its will.

Computing power sharing layer: When a computing unit is idle, it reports its computing power to the scheduler and communicates with an assigned requestor to get task’s data, execute the task, and send the result back. As we mentioned earlier, there are one-off resources and constant resources from the requester. Constant resources are sent to the computing unit first and one-off resource will be sent next. For the needs of task re-allocation strategy, the com-puting units will send its current system performance indi-cators along with result to the requester.

Monitor layer: Computing unit monitors its current exe-cuting tasks. If the job is a sub-task, the data transmission of ‘‘Pre-fetching Process” will start which will pre-fetch the necessary data of the next sub-task and send back the pre-vious processing result to the requestor.

Communication layer: Computing units discover and connect the system scheduler through broadcasting mes-sages by brokers in the communication layer.

4.4. NB-CPS event ﬂow diagram

There are many processes in NB-CPS and they could be described by the basic event ﬂow diagram as shown in Fig. 12. All the processes must start with connection estab-lishment, scheduler initialization, task allocation, some crashing handling events and end with task re-allocation. Since NaradaBrokering provides broker crashing handling

COMPUTING UNIT (1)

REQUESTOR(2)

SearchScheduler(task quantity, priority) x

SearchScheduler x

REQUESTOR(3)

SearchScheduler(task quantity, priority) x

SchedulerResponse

SearchScheduler(task quantity, priority)

SearchRequestor(QoS)

REQUESTOR(4)

SearchScheduler(task quantity, priority)

SchedulerResponse

COMPUTING UNIT (5)

SearchScheduler

SchedulerResponse

SearchRequestor(QoS)

Wait For Response > Threshold

Fig. 13. The scheduler initialization process.

SCHEDULER REQUESTOR COMPUTING UNIT (6) R epor t (1-a ) R egist er ta sks (2 ) As sign a re qu es_to r (1 -b₎ A sk_f or a re que sto_r (3) Ask for a task

(4) Send the task (5) Send the result

(11)

mechanisms, the broker crashing events will not be dis-cussed here. The events mentioned in Fig. 12 will be explained next.

Connection establishment

Each node has to enter NB-CPS network and connect to a broker with a unique identification number. To play the roles in the system, each node has to subscribe to related templates. For example, a requestor has to subscribe to the template ‘‘ToRequestor”. When a node wants to broad-cast its message to all requestors in the system, it can simply publish its message through the template ‘‘ToRequestor”. Other than that, each node has to subscribe to the template ‘‘To/X” where X is represented as its unique identification number. When a node wants to send a message to a node with the identification number ‘‘1111”, it can simply publish its message through the template ‘‘To/1111”.

Scheduler initialization

The scheduler initialization process and interacting mechanism shown in Fig. 13will explain how a requester knows there is already an existing scheduler and interacts with other roles. Supposing computing unit ‘‘1” enters the system ﬁrst (the number ‘‘1” represents its identiﬁcation number), it searches for the scheduler and reports its avail-ability by broadcasting through the template ‘‘ToSchedul-er”. At that moment, there is no scheduler in the system, thus, computing unit (1) will not be responded.

SCHEDULER REQUESTOR(1) REQUESTOR(2) COMPUTING UNIT

SearchScheduler SchedulerResponse x SearchScheduler SearchScheduler x SchedulerResponse SchedulerResponse RegisterTask(Quantity, Priority)

ResendTaskReports(Aged Task Data)

SearchRequestor

Wait For Response > Threshold

Fig. 15. The process of handling the scheduler missing event.

Get History Record <system state/task info/time>

Find High/Low Level of Execution Time <K-Mean Clustering Algorithm >

Predict Execution Time Find Loading Level

Predict Remaining Time

Remaining Time > Execution Time

Re-allocate Task YES

Idle Computing Unit <system state > Find the Critical Feature

Find High /Low Level of Critical Feature by SVM for

K-Mean Clustering Algorithm

Busy Computing Unit <start time /system state >

Find Loading Level

NO

Fig. 16. The K-means prediction process.

Table 2

The requirements of system environment Characters Environment

Requestor Windows XP/2000/2003

Scheduler Sun JAVA Runtime Environment1.4–1.5.11 Computing Unit Microsoft Access 2000

Broker Windows XP/2000/2003

Sun JAVA Runtime Environment1.4 – 1.5.11 NaradaBrokering 1.2.1

(12)

Next, requestor ‘‘2” comes into the system and searches for the scheduler by broadcasting its computing request through the template ‘‘ToScheduler”. Due to the absence of the scheduler, requestor (2) will become a scheduler after a pre-deﬁned waiting time and broadcast the role entitle-ment in the system (computing unit ‘‘1” and requestor ‘‘3”) through the templates ‘‘ToComputingUnit” and ‘‘ToRequestor”. After receiving this announcement in the system, requestors will resend their computing requirements to the scheduler and computing unit will resend their avail-ability to the scheduler too. Requestor ‘‘4” and computing unit ‘‘5” who comes into the system subsequently will be responded by the scheduler after their requests are sent.

Weighted priority based task allocation process

When there are un-allocated tasks, a message of com-puting unit’s availability will trigger the task allocation process as shown in Fig. 14 for scheduler. The scheduler selects tasks according to their priorities. For example, sup-pose there are requestor A and B with priorities 5 and 1. When the task selection begins, A’s task will have 5 times of possibility to be selected than B’s task. After selecting a task, the identification number of the task’s owner will be sent to the computing unit. If the computing unit haven’t executes the requestor’s task before, constant resources of the task will be sent first. In NB-CPS system, the file resources will be transmitted through peer-to-peer format and computing units will make reservations to requestors through Pub/Sub model before the P2P file transmission.

Fig. 19. The user interface of system scheduler monitoring computing units. Fig. 17. The user interface of a system requestor.

(13)

The missing scheduler event handling process

To avoid the catastrophe of the entire system if the sched-uler is missing unexpectedly, there is a missing schedsched-uler event handling process in NB-CPS system and an example is shown inFig. 15. Each requestor sends a testing message to the scheduler periodically. If the scheduler doesn’t respond within a certain time period (ﬁve times longer than the time interval of two testing messages), the scheduler will be considered missing and will be replaced by a requestor who discovers the event ﬁrst (requestor ‘‘2” in Fig. 15). The new scheduler will broadcast its new position to other nodes and associated required information for the new scheduler will be sent from the existing requesters.

The missing requestor event handling process

To ensure all requestors are still alive, scheduler moni-tors requesmoni-tors by recording the timestamp of testing mes-sages sent by requestors. If the interval of two testing messages are larger then a threshold, the scheduler will

consider that the requestor already left and its remaining tasks will not be allocated any more.

The missing computing unit event handling process When a requestor sends a task to a computing unit, this timestamp will be recorded. When all other tasks of this requestor are ﬁnished, and the execution time of the unﬁn-ished task reaches a threshold (5 times than the average execution time), the tasks will be reported to the scheduler and will be reassigned to other available computing units. The task re-allocation strategy using K-means prediction

process

While all subtasks are allocated, the task re-allocation process will be triggered if there are available computing units and busy ones. The goal of task re-allocation is to let idle computing units with higher performance acquire tasks which are executed by computing units with less com-puting powers to reduce the total execution time. The scheduler will first gather all unfinished subtasks’ informa-tion including subtask’s name, subtask’s serial numbers, identification numbers of computing units, starting time and system status from the requestor. When an idle com-puting unit asks for new sub-tasks, the K-means prediction process will be triggered to re-allocate unfinished tasks as shown inFig. 16.

The scheduler ﬁrst analyzes historical data of ﬁnished subtasks to obtain the critical performance indicators of the tasks by calculating the correlation between the values of each performance indicator from each execution time. The K-means prediction process next divides the values

Fig. 20. The user interface of a computing unit.

(14)

of the critical performance indicators of historical data into two groups: high loading level and low loading level. From these two loading levels, the execution time of these two levels will be calculated accordingly.

Since the task reallocation process is triggered, the scheduler will estimate the system loading level of the idle computing unit and predict its execution time. If the remaining time of an unﬁnished task is longer than the pre-dicted execution time of an idle computing unit, the task will be re-allocated. While multiple jobs could be reallo-cated, high priority subtasks will be reallocated ﬁrst.

5. System implementation and performance analysis The NB-CPS architecture is implemented in this study with open source software and the system requirement is shown inTable 2. The current system has completed the testing on JRE 1.4 to JRE 1.5.11.Fig. 17 shows the user interface of the requestor which includes the information of a broker IP address, priority of its task, resource’s ﬁle name of subtasks, and so on. Multiple executable ﬁles of a subtask are also acceptable in the system in order to per-form the sequential tasks.Fig. 18shows the user interface

Fig. 22. (a) The execution time of 4MB file size with/without network flow mechanism for NB-CPS. (b) The execution time of 130 MB file size with/ without network flow mechanism for NB-CPS.

(15)

of the scheduler who is monitoring the operation of the requestors with the progress status of each task. Fig. 19 shows the user interface of the scheduler who is monitoring the computing units with the current status. The user inter-face of a computing unit is shown inFig. 20with its current working state.

5.1. Distributed computing coordination experiment with a single requestor

In order to test the performance of the NB-CPS system, a digital watermarking algorithm with wavelet ﬁlter bank

selection is performed (Tsai, 2004). The DWT-based digital watermark algorithm consists of decomposition, embed-ding, reconstruction, and detection procedures (Cox, Kilian, Leighton, & Shamoon, 1997; Tsai, 2004). Through the com-parison of original watermark and embedded watermark from the attacked image, a similarity function based on cor-relation statistics is calculated and the authority of the digital image can be veriﬁed. By using the algorithm, digital copy-right properties can be well protected and the ownership information can be preserved under attacks.

The wavelet filter-evaluation algorithm test total 76,177 filters which are grouped into several tasks with 100 filters

Fig. 24. The accumulated number of completed tasks with diﬀerent priorities.

Fig. 23. (a) The network throughput demonstration without flow control mechanism. (b) The network throughput demonstration with flow control mechanism where red represents downloading flow, green represents uploading flow and yellow represents the flow of downloading and uploading at the same time. (For interpretation of the references in colour in this figure legend, the reader is referred to the web version of this article.)

(16)

in each task. Each task is to extract the watermark from a watermarked raw image and JPEG2000 image by different filters and to calculate the correlation coefficient of two images. The task is highly dependent on the computing power and filters can be operated in parallel, thus, this experiment is suitable for the verification of the system coordination.

The computing units we choose are with similar perfor-mance (Intel Pentium 4 3.2 GHz and 512 MB DDRII) and they are all dedicated for the computing assignment. After steadily increasing the number of computing units in NB-CPS system, the total execution time is decreasing gradu-ally. By comparing the NB-CPS with WS-CPS, we can ﬁnd the NB-CPS has better system eﬃciency from Fig. 21 by

about 10–15% improvement in reducing the total process-ing time.

5.2. The large ﬁle transmission experiment

As we mentioned previously that the WS-CPS exists the issue of large file transmission, we have proposed a network flow control mechanism for NB-CPS to deal with this prob-lem and the simulation is performed under a network where the bandwidth is up to 100 MB/s. The operation is straight-forward that the requestor will send 20 files of equal size to the available computing units and the computing units will directly send back the files to the requester. The transmis-sion time versus the number of computing unit is shown in Fig. 22a, where the file size is 4 MB each transmitted through NB-CPS. Apparently, the processing time is decreasing while the number of computer units increases since the requestor could send files to available computing units while there are more available computing units. In Fig. 22a, the execution time for NB-CPS with or without flow control is almost identical which demonstrates the flow control does not incur any delay for system performance.

However, the simulation with the transmitted file size of 130 MB will have significantly different results as shown in Fig. 22b. Since large files consume more network band-width and the I/O time of the computing unit, the network transmission is congested when the amount of the comput-ing units reaches the threshold (‘‘3”, in this case) and the execution time increases simultaneously. Fig. 23a demon-strates the network bandwidth utilization is not in peak value while there is no flow control mechanism. On the other hand, the network flow control mechanism helps the system with reliable flow transmission as shown in

Fig. 25. Diﬀerent performance indicators correlated with execution time.

(17)

Fig. 22b where the transmission time is almost the same when the computing units increase and utilize the network bandwidth efficiently which is demonstrated in Fig. 23b and the network flow almost remains at the peak speed. Therefore, the network flow control mechanism could achieve reliable quality of service for NB-CPS.

5.3. The veriﬁcation of weighted priority based task allocation

While there are multiple requestors with different prior-ity tasks, the simulation involves 3 requestors with three different priorities ‘‘1”, ‘‘2” and ‘‘3” tasks for the filter bank selection experiment with 100 subtasks each. Priority 3’s task has 3 times higher probability than 1’s task and 1.5 higher probability than 2’s task to be selected in the weighted priority based task allocation. As shown in Fig. 24. The accumulated number of completed tasks is increasing as times grows and the priority 3’s tasks com-pleted first, the priority 2’s job comcom-pleted next. However, NB-CPS system can manage multiple requesters with mul-tiple priority tasks and the weighted priority bask task allo-cation could guarantee there is no starvation situation for low priority tasks fromFig. 24.

5.4. The veriﬁcation of K-means prediction process

As noted in the previous section, NB-CPS collects the historical executing information and predicts the execution time of the task using K-means prediction process. It is essential to have accurate ﬁndings of the critical performance indicator for the task. In Fig. 25, it shows the correlation values of the system performance indicators when a computing unit executes the ﬁlter bank selection

experiment under a dynamic environment. From the plot, ‘‘CPU usage” has the highest correlation value at 0.76 than CPU speed and RAM usage which should become the crit-ical performance indicator of the task.

To verify the K-means prediction process, 5 computers with diﬀerent capacity randomly running Super PI (Super, 2005) to simulate the dynamic environment is performed. Based on the SVM classiﬁcation, two clustering has been grouped for each computing unit and the average execution time at the low and high loading are shown respectively in Fig. 26. Use this data for the K-means prediction process, we have performed 1000 times of the task allocation for NB-CPS and the average prediction error is calculated by the Eq.(3)

Average prediction error¼ P5 j¼1 P200 i¼1 jTPij–TRijj TRij 1000 ð3Þ

where TPijis the predicted processing time of task i at com-puting unit j and TRijis the actual processing time of task i at computing unit j. The average prediction error is only 11.0% which shows that ‘‘K-means prediction process” is an eﬃcient approach of predicting the execution time.

It is also import to compare the performance between the ‘‘K-means prediction process” and the CC-FGDM of WS-CPS (Wang et al., 2007). CCC-FGDM is a fuzzy group decision-making technique which assigns each sys-tem performance indicator the same weighting which is generated by analyzing the historical records through regression analysis. The objective of this experiment is to ﬁnd out whether ‘‘K-means prediction process” or CC-FGDM can eﬀectively predict which computer has the min-imum execution time. Same simulation environment for 5 computing units under dynamic environment is performed and the wrong prediction rate for K-means predication

(18)

process is 1.77% but CC-FGDM is 22.8% for reallocation. Therefore, the data demonstrates K-means prediction could have suﬃcient prediction capability. The overall pro-cessing time comparison is shown inFig. 27. While the task number is increasing, the execution time is increasing for both schemes but NB-CPS is always with shorter process-ing time than CC-FGDM of WS-CPS approach by about 10% improvement. Therefore, we can conclude the NB-CPS design has superior performance than WS-NB-CPS tech-niques under dynamic environment.

6. Conclusion

In this article, NaradaBrokering Based Computing Power Service (NB-CPS) is proposed which is not only resolve the issues in WS-CPS such as system resilience, fault tolerance, efficiency of job scheduling and the instability in congested network environment, but also efficiently execute computa-tion intensive works. NB-CPS sufficiently utilizes the P2P grids integration with the bandwidth flow control mecha-nism by FCFS reservation strategy, weighted priority based task selection, SVM classification and K-means prediction process to resolve the task allocation strategy. Experiments confirmed that NB-CPS has superior effectiveness in single requestor situation, multiple requestors situation, interfered environment, and in high traffic environment.

Acknowledgements

This work was supported by the National Science Coun-cil in Taiwan, Republic of China, under Grant NSC95-2416-H009-027 and NSC96-2416-H009-015.

References

Banavar, G., et al. (1999). An eﬃcient multicast protocol for content-based publish-subscribe systems. In Proceedings of the IEEE interna-tional conference on distributed computing systems.

Casanova, H. (2002). Distributed computing research issues in grid comput-ing. ACM SIGACT News Distributed Computing, Column, 8, 50–70.

Chang, C.-C., & Lin, C.-J. (2001). LIBSVM: A library for support vector machines, software. Available at<http://www.csie.ntu.edu.tw/cjlin/ libsvm>.

Cox, I. J., Kilian, J., Leighton, F. T., & Shamoon, T. (1997). Secure spread spectrum watermarking for multimedia. IEEE Transactions on Image Processing, 6(12), 1673–1687.

Erradi, A., et al. (2006). Diﬀerential QoS support in Web Services Management. In International conference of Web services, ICWS ‘06 (pp. 781–788).

Foster, I. (2002). The grid: A new infrastructure for 21st century science. Physics Today, 55(2), 42.

Foster, I., Kesselman, C., & Tuecke, S. (2001). The anatomy of the grid: Enabling scalable virtual organization. International Journal of High Performance Computing Applications, 15(3).

Fox, G., et al. (2003). NaradaBrokering: A middleware framework and architecture for enabling durable peer-to-peer grids. In Proceedings of the international middleware conference.

Linpack (2006).<http://www.netlib.org/linpack/>.

MacQueen, J. B. (1967). Some methods for classiﬁcation and Analysis of Multivariate Observations. In Proceedings of 5-th Berkeley Symposium on Mathematical Statistics and Probability (Vol. 1, pp. 281–297). Berkeley: University of California Press.

Plank, J. S. et al. (2001). Processor allocation and checkpoint interval selection in cluster computing systems. Journal of Parallel and Distributed Computing, 61(11), 1570–1590.

Segall, B., Arnold, D. (1997). Elvin has left the building: A publish/ subscribe notiﬁcation service with quenching. In Proceedings AUUG97 (pp. 243–255).

Super, P. I. (2005). Available at: <http://www.super-computing.org/> (Accessed December, 2005).

The NaradaBrokering Project (2006). Indiana University. <http:// www.naradabrokering.org/>.

Tsai, M. J. (2004). Filter bank selection for the ownership veriﬁcation of wavelet based digital image watermarking. In Proceedings of the 2004 international conference on image processing (Vol. 5, pp. 3415– 3418).

Tsai, M. J. et al. (2006). A collaborated computing system by web services based P2P architecture, computer supported cooperative work in design II. Lecture Notes in Computer Science, 3865, 194–204. Tsai, M. J., & Wang, C. S. (2008). A computing coordination based fuzzy

group decision making (CC-FGDM) for web service oriented archi-tecture. Expert Systems With Applications, 34(4), 2921–2936. Viktor, S., Eide, E., et al. (2003). Extending content-based publish/

subscribe systems with multicast support, Technical Report 2003–03, Simula Research Laboratory, pp. 1 (July 2003).

Wang, S. C. et al. (2007). A hybrid load balancing policy underlying grid computing environment. Computer Standards and Interfaces, 29, 161–173.