The Architecture of SIG Computing Environment and Its ... - CiteSeerX

1 downloads 0 Views 111KB Size Report
SIG computing environment is mainly composed of GT [7], OpenPBS, and. Condor/Condor-G [8]. OpenPBS and Condor are the local resources schedulers, each.
The Architecture of SIG Computing Environment and Its Application to Image Processing 1

1

1

1

2

Chunhui Yang , Deke Guo , Yan Ren , Xueshan Luo , Jinfeng Men 1

School of Information Systems and Management, National University of Defense Technology, Changsha, China, 410073 [email protected], [email protected], [email protected] 2 Key Lab of CFC, National University of Defense Technology

Abstract. Spatial Information Grid (SIG) is a project of applying grid technology to share and integrate spatial data resources, information processing resources, equipment resources, and knowledge resources. SIG computing environment aims to apply the concept of SIG to share hybrid computing resources for processing remote sensing (RS) images. RS image processing is a data-intensive computing problem, and it adapts to be processed according data parallel computing model. In this paper, we discuss the architecture of SIG computing environment, which can provide a powerful computing infrastructure used to process RS image cooperatively. In order to achieve high performance, we propose a model of the image division. From the relation among the processing time, the communication latency, and the transferring ratio, we can achieve some useful conclusions to determine the strategy of the image division. Furthermore, we can discover two optimal division strategies through comparing the experimental results with those useful conclusions.

1. Introduction With the rapid development of RS technology and the increasing complexity of computation, more and more RS images make the processing of them in a single PC almost impossible [1]. Now we can utilize the idle resources in the Internet by grid computing technology to finish many tasks which are data intensive or computing intensive. Till now, there have been some successful cases, such as EU dataGrid [2], SETI@home [3], ChinaGrid [4], VEGA [5], and so on. The task of processing the RS image just has the properties of data intensive and computing intensive, and then it adapts to parallel processing. Dividing a computing task into smaller computing tasks and assigning them to different processors for parallel execution are the two key steps in the design of parallel algorithms [6]. In this paper, we focus on the RS image division, which is the key step of decomposing the computing task. In the division model, we discuss the situations respectively whether the idle time between communication and computing on a host exists or not. Furthermore, we get two optimal division strategies according to the trend of the processing time with the image size and the relation among processing time, communication latency, and transferring time.

The remainder of the paper is organized as follows. The next section introduces the architecture of SIG computing environment. Section 3 focuses on the computing model of dividing image. And we validate the conclusions by the experimental results. Finally, we draw the conclusion and give out the future work in section 4.

2. The Architecture of SIG Computing Environment SIG computing environment is mainly composed of GT [7], OpenPBS, and Condor/Condor-G [8]. OpenPBS and Condor are the local resources schedulers, each of which is responsible for managing a cluster. And the cluster’s worknodes are the foundation of SIG computing environment, whose responsibilities are providing computing power. Condor-G is the upper level. Users directly contact Condor-G to submit the jobs and monitor the status of jobs. GT establishes a bridge between batch systems and Condor-G. It supplies many essential services and protocols for grid computing, such as data management, resource allocation and management, security service, and information services [9][10]. The architecture of SIG computing environment is shown in Fig.1 [11][12]. C luster Subm it

C ond or-G Sched uler

JobM anager

U ser

PB S Server

gatekeeper

Persisten t Jo b Q ueue

JobM anager

PB S Sch eduler Fro nt-end

W orknod es

C ond or-G 1

G ridM anager

C luster

gatekeeper

JobM anager G A SS Server gatekeeper

C on dor

GRAM JobM anager

Job Subm ission M achine

W orknod es Fro nt -end

Fig. 1. The grid environment architecture.

3. The Strategy of the Image Division It is critical that scheduling the tasks of processing the small image pieces in SIG computing environment for achieving high performance. In our scenario, we predetermine the size of each image piece. Then the parallel tasks are fixed-sized tasks. At the same time, there are no task synchronizations and no inter-task communications. Therefore, the scheduling problem is challenging due to the communication and transferring time involved when starting tasks [13]. From [14], we know that in any optimal solution all worknodes should participate in the computation. In SIG computing environment, the number of worknodes is predetermined. Therefore, we should divide the image into pieces which are the multiple

of the number of worknodes. Now the question is how many rounds we should make each worknode executing the tasks.

3.1 The Division Model and Some Conclusions In the model, we suppose there are N worknodes, which are homogeneous platforms. At the same time, all of them are identical. Hence, the execution time of job is the same, independent of which worknode the job is place on [15]. For communications, the one-port model is used: the master can only communicate with a single worknode at a given time-step. At the same time, we fix the communication latency to simplify the model. Before submitting a job to worknode, the master needs communication with worknodes and transfers the input data into a worknode. Moreover, the communication and computing can not be overhead, i.e. each node can be assigned a new job only after it has finished the previous job. According to whether the idle time between communication and computing on a host exists or not, we can divide the division strategy into two situations. Round2

R ound 1 S / N *m *B N ode1 N ode2

id le tim e f(S / m *N )

tla t

……

N odeN tim e

Fig. 2. There is idle time between communication and computing in situation A. R o u n d1

Round2

S /N * m *B f(S / m *N )

N ode 1 N ode 2

tla t

……

N o d eN

tim e

Fig. 3. There is no idle time between communication and computing in situation B.

Situation A: As shown in Fig.2, there is idle time between communication and computing, that is: S S (1) f(

m N

)  (tLat



m N B

) ( N 1 )

Where S is the size of input data; m is the number of round, and it is a natural number; N is the number of worknodes; tLat is a fixed overhead for starting a job, incurred by the master to initiate a job to a worknode [14]; B is the data transfer rate f (

S

)

to worknode; and m N is the time of processing each image piece, which is the function of the image size. In this situation, the whole processing time is:

(1) T wm ( tLat 

S S ) m N f ( ) m N B m N

(2)

In this situation, there is no idle time between two communications, then the whole time is the sum of the time of all communication and transferring adding the computing time of a piece. So, in situation A, the best division strategy is: S S (3) (1) minT

wm

( tLat  ) m  N f ( ) m N B m N

Where m is a natural number. And equation (3) is subject to equation (1). Situation B: As show in Fig.3, the processing time of each piece should be equal to or longer than the time of all other worknodes receiving the input data in a round. That is: S S (4) f( )  (tLat  ) ( N 1 ) m N

m N B

In this situation, the whole processing time is: S S (5) (2) T wm ( tLat  ) ( N m 1) m  f( ) m N B m N In this situation, there are no idle time between communication and computing. Therefore, the whole processing time is equal to the whole working time of worknode N add the communication and transferring time of other (N-1) worknodes in round 1. In this situation, the best division strategy will make minT

(2) wm

(2) T wm

into the minimum, i.e. (6)

S S ( tLat  ) ( N m 1 ) m  f( ) m N B m N

Where m is a natural number. And equation (6) is subject to equation (4). (1) Therefore, the final division strategy is the smaller one of minT wm and

(2) minTwm

.

N pieces. Then the image will be divided into m

We can get the best division strategy by calculating according to the above equations. But after analyzing the model, we can get some simple conclusions. At first, we suppose the three trends of processing time with the image size are case A, B and C, each means that S (m 1) f( ) (m 1) N

m  f(

S ) m N

is greater than, equal to, and less than

.

In situation A, it is easy to decide the value of m, we define the T wm : S S (1) (1) (1 ) Twm Twm Tw(m N  tLat m  f( ) ( m 1)  f( ) 1) m N (m 1) N (1)

(7)

T (1) 0

wm If , the image should be divided unceasingly. Thus, if the function of processing time belongs to the case B or C, we should not continue dividing the image. Or else, we should continue dividing the image.

T (2)

wm In situation B, we define the : 1 ( N  1 ) S S S (2) (2 ) Twm Twm Tw(2)(m1)   tLatm f( ) (m 1)  f( ) m (m 1) N  B m N (m 1)  N

(8)

Where tLat is usually small and the first item is always positive. Therefore, in equation (8), if f(x))belongs to the case A or B, we should continue dividing the image

until it is restricted by the condition of equation (6). When f(x))belongs to the case C, it is difficult to make decision of division. We can only get the decision by calculation. In our application, the time of division and makeup is the function of image size and not changing with the number of pieces, and the reduction of the whole time is only connected with the function of edge detecting.

3.2 Experimental Results As for our application, the whole system is a mode of master and worknode. In our experiment, we apply the PBS client as the worknode and there are 4 worknodes. The parameters are as follows: N =4; tLat =3.0s; B =10.25Mb/s; S =467.0Mb. F(m) f(

S

) T comm

( tLat 

S

) ( N 1)

m N , m N B Suppose , the relation of them in our application can be contracted as Fig.4. As shown in Fig.4, when m 2 ,

F(m) T comm m f(

,

it

belongs

to

situation

B.

At

the

same

time,

S S )  (m 1) f( ) m N (m 1) N

, it accords with the condition of case A. Therefore, according to the above conclusion, m =2 is the best strategy in situation B. F(m)Tcomm When m 3, , it belongs to situation A, and it also belongs to case C. Thus, m =3 can get the shortest executing time in situation A. 80 70 60 50 40 30 20 10 0

150 F(m) Tcomm

100 50

m=1

m=2

m=3

m=4

Fig. 4. The relationship of F(m) and Tcomm.

0 4

8

12

16

Fig. 5. Different divisions lead to different processing time.

As shown in Fig.5, when the image is divided into 8 pieces, it is the best division strategy of situation B. And 12 pieces is best for situation A. Therefore, the final strategy is dividing the image into 8 pieces. The experimental results show that the division strategy is optimal.

4. Conclusion and Future Work This paper presents SIG computing environment, which is based on the grid middleware Globus Toolkit and utilizes the computing resources with hybrid structure. We apply the SIG environment to the processing of RS image. We propose the model of the image division, which can compute the most appropriate image pieces and

make the processing time short. And experimental results show that the model and the conclusions are logical. In the future, we will develop the web portal by using the API of Condor-G and enclose the whole workflow of image processing into a service, which makes it more easily to apply. On the other hand, the study of load balance is an important part of future work.

Acknowledgement This work is supported by the National High Technology Research and Development Program of China under grant No.2002AA104220, 2002AA131010, and 2002AA134010.

References 1. H. Zhou, X. Yang, and Y. Tang et al, Research and Implementation of Grid-Enabled Parallel Algorithm of Geometric Correction in ChinaGrid, The 3rd International workshop on Grid and Cooperative Computing (GCC2004), Wuhan, 2004, pp. 704-711. 2. http://eu-datagrid.web.cern.ch/eu-datagrid. 3. http://setiathome.ssl.berkeley.edu. 4. http://www.chinagrid.edu.cn. 5. Z. Xu and W. Li, Research on VEGA Grid Architecture, Journal of Computing Research and Development, Vol.39, No.8, 2002, pp. 923-929. 6. A. Grama, A. Gupta, and G. Karypis et al, Introduction to Parallel Computing, Second Edition, Addison Wesley, 2003. 7. http://www.globus.org. 8. http://www.cs.wisc.edu/condor. 9. I. Foster and C. Kesselman, The Grid2: Blueprint for a New Computing Infrastructure, Morgan Kaufmann Publishers, 2004. 10. S. Chen, W. Zhang, and F. Ma et al, The Design of a Grid Computing System for Drug Discovery and Design, The 3rd International workshop on Grid and Cooperative Computing (GCC2004), Wuhan, 2004, pp. 799-802. 11. J. Frey, T. Tannenbaum, and I. Foster et al, Condor-G: A Computation Management Agent for Multi-Institutional Grids, Proceedings of the Tenth IEEE Symposium on High Performance Distributed Computing (HPDC10) San Francisco, California, 2001, pp.55-64. 12. A. Waananen, M. Ellert, and A. Konstantinov et al, An overview of an Architecture Proposal for a High Energy Physics Grid, Proc. Applied Parallel Computing (PARA 2002), LNCS 2367, Springer-Verlag, 2002, pp. 76 -86. 13. O. Beaumount, H. Casanova , and A. Legrand et al, Scheduling Divisible Loads on Star and Tree Networks: Results and Open Problems, IEEE Transactions on Parallel and Distributed Systems, Vol.16, No.3, 2005, pp. 207-218. 14. Y. Yang, H. Casanova, Multi-round algorithm for scheduling divisible workload applications: analysis and experimental evaluation, Tech. Rep. CS2002-0721, Department of Computer Science and Engineering, University of California, San Diego, 2002. 15. T. Hsu, J. C. Lee, and D. R. Lopez et al, Task Allocation on a Network of Processors, IEEE Transactions on Computers, Vol.49, No.12, 2000, pp. 1339-1353.

Suggest Documents