4. Comparisons
The matching pursuit decomposition of seismic data is performed trace by trace. Actually, for a seismic profile or a 3D seismic data volume, the traces can be directly assigned to the compute nodes of parallel computer systems on average to implement decompositions in parallel. We use a seismic profile as shown in Figure 9 to compare this coarse-grained parallel strategy and our parallel algorithm of matching pursuit. The operating environment of these two parallel programs is still an Intel multiprocessor computer whose main frequency of processors is 2.40GHz.
First, those 800 traces are assigned to 1, 2, 4, 8, 16 and 32 processors on average to decompose in parallel, respectively. Then we implement our parallel algorithm to the traces by setting 1, 2, 4, 8, 16, 32 processors, respectively, and searching only one optimal wavelet by every processor. In these two parallel methods, we define the same threshold value to end the decompositions for fair comparisons.
After all the implementations are completed, we calculate the speedup and efficiency of these two methods for different numbers of processors, respectively, as shown in Figure 10. The speedup of the coarse-grained parallel strategy is apparently lower than that of our parallel algorithm of matching pursuit, and the difference between these two is increased with the increasing number of processors. Meanwhile, the overall efficiency of the coarse-grained parallel strategy is also much lower than that of our parallel algorithm. That’s because the number of iterations and optimal Morlet wavelets of seismic traces are unknown before the decomposition according to the matching pursuit algorithm. The coarse-grained parallel strategy easily leads to uneven jobs for compute nodes even though we assign the same number of traces to every compute node, affecting the overall efficiency of the parallel computing.
