| Article Index |
|---|
| Application of cluster computing for seismic imaging in the energy industry |
| Page 2 |
| Page 3 |
| Page 4 |
| Page 5 |
| All Pages |
Performance Analysis
Even though the developed codes for ?-x depth migration (with and without parallel I/O) are portable across various platforms, most of the development was done on PARAM 10000. The PARAM 10000 system has 40 SUN E450 compute nodes, each with 4 processors @300MHz. Out of 40 nodes 4 nodes are network file servers with 1GB RAM and 512K cache. High-speed network such as fast Ethernet with peak bandwidth 100MB/s connects the nodes. Following tables show the comparison between execution time for ?-x depth migration algorithms with and without parallel I/O for various small and large sized data sets.
2-D Depth Imaging
Size of FFT Data: 2.66 Mb
Size of velocity model: 1.824 Mb
Total number of frequencies used: 222
Execution time with 16 processors
Without parallel I/O: 57.8 seconds
With parallel I/O: 37.6 seconds
3-D Depth Imaging, Dataset 1
Size of FFT Data: 60 Mb
Size of velocity model: 75 Mb
Total number of frequencies used: 420
Execution time with 24 processors
Without parallel I/O: 5765 seconds
With parallel I/O: 4331 seconds
Execution time with 64 processors
Without parallel I/O: 2301 seconds
With parallel I/O: 1645 seconds
3-D Depth Imaging, Dataset 2
Size of FFT Data: 1.3 Gb
Size of velocity model: 1.2 Gb
Total number of frequencies used: 256
Execution time with 64 processors
Without parallel I/O: 43500 seconds
With parallel I/O: 28370 seconds
Discussion of Parallel I/O Performance
The execution time charts given above show that with parallel I/O implementation the seismic imaging algorithm takes almost 30% less time compared to the standard I/O implementation. The reason for this improvement is the change in communication that is possible with the support of parallel I/O. In the case of parallel I/O, the communication involved in the reading and distributing the data among processors can be changed to just reading the data in parallel by the processors without any communication involved. The communication comes at the end of the algorithm, when the final imaging takes place.
In the seismic industry, where the amount of data that needs to be processed is often measured by the number of tapes, which amount to hundreds of gigabytes and now terabytes, the improvement of making efficient use of the I/O subsystem becomes increasingly apparent. A 10% to 20% improvement in runtime would amount to saving of millions of dollars in processing time. The case study discussed above highlights a step in this direction.
References
1. Claerbout. J., Imaging the Earth's Interior, 1985, Blackwell Publications
2. S. Phadke, D. Bharadwaj and S. Yemeni, Wave equation based migration and modeling algorithms on parallel computers., Proc. of SPG , 1998.
