| 摘要 |
<p>This invention provides a method for inputting seismic trace data to a seismic processing program running on a massively parallel processor (MPP) and for balancing the computational workload of individual processors. The MPP, consisting of a plurality of individual processing nodes connected by a data and control communications network, is treated as a bipartite system of 'operator' nodes and 'input' nodes. Opertor nodes apply mathematical operators to transform seismic data. Groups of seismic traces are loaded into the input nodes. The input nodes examine the data to determine which regions of the seismic survey are covered by the traces and how many traces cover each bin. The input nodes then broadcast to all operator nodes a compact description of the trace coverage. The operator nodes then retrieve traces from the input nodes, using efficient data transfer mechanisms of the data network. The input nodes can be seen to serve as a distributed, content-addressable data cache. All nodes use the information in each broadcast to determine how many bins will be assigned to each 'operator' node that will process the prestack data to produce poststack data. The bin assignment is designed to equalize the number of traces that each opertor node processes. Thus, the invention provides a means for maximizing the efficiency with which an MPP can perform seismic migration processing, and is applicable to wide variety of MPPs and processing algorithms.</p> |
