3.1 Matrix Multiplication Results
The matrix multiplication was run with forking of different umbers of tasks to demonstrate the speedup. The problem sizes were 256X256, 512X512, 768 X768, 1024X1024, and 1280X1280 in our experiments. It is well known, the speedup can be defined as ts / tp, where ts is the execution time using serial program, and tp is the execution time using multiprocessor. The execution time o dual2 (2 CPUs),dual2 ~3 (4 CPUs),dual2 ~4 (6 CPUs),dual2 ~5(8CPUs), and dual2 ~9 (16 CPUs), were listed in Figure 5, respectively.The corresponding speedup of different problem size by varying the umber of slave programs were shown in Figure 6.Since matrix multiplication was uniform workload application, the highest speedup was obtained about 10.89 (1280 ?1280)by using our SMP cluster with 16 processors. We also found that the speedups were closed when creating two slave programs o one dual processor machine a d two slaves program on two SMPs respectively.
Figure 5: Execution time (sec.) of SMP cluster with different number of tasks (slave programs).
3.2 PVMPOV Cluster Benchmark Results
Pov-ray is a multi-platform, freeware ray tracer [4]. Many people have modified its source code to produce special "unofficial" versions. One of these unofficial versions is PVMPOV, which enables POVray to run o a Linux cluster.
Figure 6: Speedup of SMP cluster with different number of tasks (slave programs).
With the cluster configured, runs the following command to begin the ray tracing and generates the image file as show in Figure 7:
./pvmpov +iskyvase.pov +w640 +h480 +FT +v1 -x -d +a0.300 -q9 -mv2.0 -b1000
-nw32 -nh32 -nt4 -L/home/gs17/pvmpov3\_1e\_1/povray31/include
This is the benchmark option command-line with the exception of the -nw and -nh switches, which are specific to PVMPOV and define the size of image each of the slaves will be working o .The -nt switch is specific to the umber of tasks will be running. For example, -nt4 will start four tasks, one for each machine. The messages on the screen should show that slaves were successfully started. When completed, PVMPOV will display the slave statistics as well as the total render time. Using single processor mode of a dual processor machine for processing 1600 X1280 image, the render time was 369 seconds. Using both CPU ' s o a single machine reduced the render time to 190 seconds. Adding the second machine with dual CPU ' s dropped the time to 99 seconds. Using out SMP cluster (16 processors) further reduced the time to 27 seconds. The execution time o dual2, dual2 ~3, dual2 ~4, dual2 ~5, and dual2 ~9, were show in Figure 8, respectively. The corresponding speedup of different problem size by varying the umber of task (option:-t) were show in Figure 9. The high speedup were gained about 1.94 (1600X1280) on dual2, and 3.73 (1600 ?1280) using both dual2 and dual3.The highest speedup was obtained about 13.67 (1600 X1280)by using our SMP cluster with 16 processors.
Figure 7: The skyvase.tga genreated form PVMPOV.
4 Conclusion and Future Work
Scalable computing clusters, ranging from a cluster of (homogeneous or heterogeneous) PCs or workstations, to SMPs, are rapidly becoming the standard platforms for high-performance and large-scale computing. It is believed
that message-passing programming is the most obvious approach to help programmer to take advantage of clustering
Figure 8: Execution time (sec.) of SMP cluster with different number of tasks (slave programs).
Figure 9: Speedup of SMP cluster with different number of tasks.
symmetric multiprocessors (SMP) parallelism. In this paper, we present the basic programming techniques by using Linux/PVM to implement a PVM program. The matrix multiplication a d parallel ray tracing problems are illustrated and the experiments are also demonstrated on our Linux SMPs cluster. The experimental results showed that the highest speedup were 10.89 and 13.67 respectively for matrix multiplication a d PVMPOV, when the umber of processors is 16, by creating 16 tasks o SMPs cluster. The results of this study will make theoretical and technical contributions to the design of a PVM program o a Linux SMP clusters for remote sensing data processing.
References
- R.Buyya,High Performance Cluster Computing: System and Architectures Vol.1,Prentice Hall PTR,NJ,1999.
- R.Buyya,High Performance Cluster Computing: Programming and Applications Vol.2,Pre ticeHallPTR,NJ, 1999.
- G.F.P . ster,In Search of Clusters Prentice Hall PTR,NJ,1998.
- http://www.haveland.com/povbench/,POVBENCH - The O . cial Home Page.
- http://www.epm.ornl.gov/pvm/,PVM - Parallel Virtual Machine.
- T.L.Sterling,J.Salmon,D.J.Backer,and D.F.Savarese,How to Build a Beowulf: A Guide to the Implementation and Application of PC Clusters 2 d Printi g,MIT Press,Cambridge,Massachusetts,USA,1999.
- B.Wilkinson a d M.Allen,Parallel Programming: Techniques and Applications Using Networked Workstations and Parallel Computers Prentice Hall PTR,NJ,1999.
