Figure shows two stacking velocity semblance panels. The top panel corresponds to the CMP location at x = 4 km, and the bottom panel to x = 4.7 km. Figure contains the same information, but is plotted in contours as opposed to raster. I have spent some effort trying to fine-tune the stacking velocity algorithm, to ``tighten'' the velocity spectrum, but perhaps some more work could be done as indicated by the residual semblance smear.
Currently, the main computational core of the CM code is running at a modest 120 Mflop/s on 8k processors. This is because the indirect memory addressing, shifting, and integer/boolean operations cumulatively require a large percentage of the cpu time compared to the low flop count. A stacking velocity analysis of 32 48-fold CMP gathers, 2,000 samples per trace, with 48 stacking velocities and 256 output time samples, requires a total of about 75 cpu seconds on 4k processors. About 30 s of that time is required to read input data from an NFS mounted disk to the front-end memory, another 10 s to transfer the data from the front-end to the CM, about 30 s to do the computations within the CM, and about 5 seconds to write the results from the CM to front-end to disk. Evidently, this process is heavily I/O bound, since I/O requires about 60% of the total run time. With a 10 Mb/s data vault or disk array, the I/O bottleneck could largely be eliminated (since the 32 gathers represent 12 Mb of data), which would speed up the effective run time by a factor of about 2.5x.