Transposing a dataset is a common seismic processing step. A given processing
step might need data in offset gathers while the data exists in common offset
sections. For source-receiver based wave
equation migration, the natural axis order is: midpoint,
offset, and depth while further processing steps need the data
in depth, offset, and midpoint. For large data volumes (significantly
beyond the CPUs memory) what looks
like a rather trivial operation can take hours to days using a single
CPU. The transpose time is dominated for large problems by disk IO.
A solution to this problem is to stream a data volume to a cluster where
the transpose can be done completely in-core. This is an effective
technique but again adds significant programming overhead.
2009-05-05